Development and Evaluation of Diversion Strategies
under Incident Response using Dynamic Traffic
Assignment System
by
Akhilendra Singh Chauhan
B.Tech. in Civil Engineering (2001)
Indian Institute of Technology, Kanpur, India
Submitted to the Department of Civil and Environmental Engineering
in partial fulfillment of the requirements for the degree of
Master of Science in Transportation
at the
MASSACHUSETTS INSTITUTE OF TECHNOLOGY
September 2003
0 Massachusetts Institute of Technology, 2003. All rights reserved.
A uthor .............................
........................
Departmeniii4{ivil and Environmental Engineering
August 9, 2003
Certified by .....................
Moshe E. Ben-Akiva
Edmund K. Turner Professor of Civil and Environmental Engineering
Thesis Supervisor
Certified by ....................
Haris N. Koutsopoulos
Associate Professor of Civil and Environmental Engineering
Northeastern University
Thesis Supervisor
Accepted by ......
-a
I
II
MASSACHUSETTS INSTITUTCh
OF TECHNOLOGY
.................
Heidi Nepf
irman, Departmental Committee on Graduate Studies
SEP 11 2003
BARKER
LIBRARIES
2
Development and Evaluation of Diversion Strategies under
Incident Response using Dynamic Traffic Assignment System
by
Akhilendra Singh Chauhan
Submitted to the Department of Civil and Environmental Engineering
on August 9, 2003, in partial fulfillment of the
requirements for the degree of
Master of Science in Transportation
Abstract
Integrating Intelligent Transportation Systems (ITS) technologies with traffic surveillance
promises to reduce the cost incurred by non-recurrent traffic congestion by improving
transportation agencies' ability to respond to and clear such incidents. Advance Traffic
Management Systems (ATMS) and Advance Traffic Information Systems (ATIS) are
examples of ITS sub-systems, which are increasingly becoming prevalent in the Traffic
Management Center (TMC) of a city in order to address various traffic management
issues. Dynamic Traffic Assignments (DTA) system can be used for dynamic traffic
management purposes, like reducing delay on major highways, improving safety,
efficiency and capacity of transportation network, and improving wide-area emergency
responses through information sharing and coordination. DynaMIT-P is a DTA-based
planning tool that operates in an offline mode, in which traffic managers investigate and
prepare standard response strategies to different traffic and incident scenarios. The
objective of this thesis is to apply DynaMIT's DTA capabilities to the Lower Westchester
County (LWC) ITS subsystem and to use calibrated system to perform illustrative
analyses of incident response strategies. First, DynaMIT-P has been calibrated in order to
estimate traffic conditions in the LWC network with precision sufficient for ITS
purposes. The results of two case studies, focusing on the evaluation of the diversion
response strategy in case of an incident on the LWC network, illustrate the functionality
and potential of the system.
Thesis Supervisor: Moshe E. Ben-Akiva
Title: Edmund K. Turner Professor
Department of Civil and Environmental Engineering
Thesis Supervisor: Haris N. Koutsopoulos
Title: Associate Professor of Civil and Environmental Engineering
Northeastern University
3
4
Acknowledgements
I would like to express my deepest regards and thanks to Prof. Moshe E. Ben-Akiva and
Prof. Haris N. Koutsopoulos for their constant support and guidance. Their constant
guidance has been invaluable to my research at MIT. In particular, I would like to thank
Haris for his constructive comments in my research at ITS lab. I was always benefited
from the many discussions I had with him. Working with him has provided me with a
deep understanding of the simulation and transportation research. He is one of the few
persons I respect the most in my life, for his vision, thinking and help.
I am grateful to Rama and Costas for their willingness to help at any hour and for
lending a patient ear to all my queries. These are the two people without whom this work
would not have materialized. Rama was extremely helpful in order to understand nittygritty details of the DynaMIT. It was almost impossible without him to understand the
various aspects of different models used in the DynaMIT. Costas was the other person
who worked along with me on this project. I learnt a lot from him, especially about the
DynaMIT software details. He was a like bridge between NYSDOT and me. His
comments and discussions have helped me in clearing several obstacles along the way of
this thesis.
My special thanks to Srini, Deepak and Kunal for their friendship, support and
help in shaping not only my research but my future career as well. Srini was very helpful
by giving his valuable comments on my thesis. He is the only person with whom I have
shared my good moments at the lab. He helped me a lot during my first year at the ITS
lab. Whenever I wanted to take some time off, Deepak and Srini frequently invited me
for dinner!!! During those dinner hours, I have shared many good moments away from
the academia. I also wish to thank Rama (Bob), Dinu, Anshiv and other members of the
gang at their home.
It is impossible to forget the moments I had with Miguel, Maya, Phani, Ale and
Michael. Miguel and I have shared light moments during the evening coffee time at the
food court. He is my best American friend, who helped to understand American slang and
culture. Maya was always very supportive with my research and course work. Phani and I
have done almost same courses at MIT, so no doubt the bond is really strong. He spent
5
hours and hours in discussing problem sets. No doubt, he is a great cook!! Ale and Mike
always used to pass some light comments, which made life easy.
Other two best friends, I made at MIT, were the members of the TDS group!!
Emmanuel, Saurabh and I had really good time and they have really lightened my
academic and research load. I would not forget my great non-academic and academic
moments with Nandu and Vimal, my roommates at MIT during the first year.
I am also grateful to my fellow people at the MIT ITS program for their help,
friendship and support. I thank Tomer for his comments and discussions and Marge for
sharing lighter moments. I also like to thank Dan, Joe, Zhili, Bhanu, Emily and Charisma
for making the lab a lively place to work. I would also like to express my thanks to all my
friends at MIT and IIT.
I like to thank administrative staff at the Center of Transportation and Logistics
(CTL), especially Leanne Russell for attending to all my requests promptly.
I thank most kindly Federal Highway Administration (FHWA) and New York
State Department of Transportation (NYSDOT) for their financial support.
Finally, I express my indebtedness to my parents, brothers and sisters for their
love, encouragement and inspiration, which have given me strength on numerous
occasions.
6
Contents
1
2
17
Introduction
1.1
B ackground ......................................................................
18
1.2
Motivation and Thesis Objectives ...........................................
21
1.3
T hesis Outline ..................................................................
22
Literature Review: Calibration Methodology and Modeling of Information
23
Dissemination
2.1
2.2
3
D TA system s ...................................................................
24
2.1.1
Simulation-based DTA systems ....................................
24
2.1.2
Model Components in DTA system ................................
27
Calibration Methodology for DTA systems ................................
30
2.2.1
Overall Calibration Approach .......................................
30
2.2.2
Demand Calibration ..................................................
31
2.2.3
Supply Calibration ...................................................
33
2.3
DTA System for Short-Term Planning Applications .....................
35
2.4
Sum mary ........................................................................
38
Development of Incident Scenarios and Evaluation of Diversion Strategies using
39
DynaMIT-P
3.1
DynaMIT: An Overview ......................................................
39
3.2
DynaMIT-P: Planning Version of DynaMIT ...............................
41
Main Components of DynaMIT-P .................................
42
3.2.1
3.3
DynaMIT-P for the Incident Scenarios and Diversion Strategies
(in general planning) ..........................................................
46
3.3.1
Modeling Day-to-Day Behavior in DynaMIT-P ..................
49
3.3.2
Modeling Within-Day Behavior in DynaMIT-P ..................
52
3.4
Representation of Incident Scenarios in DynaMIT-P .....................
52
3.5
Representation of ATIS/VMS Scenarios in DynaMIT-P .................
53
Modeling ATIS/VMS with Instantaneous Information ..........
54
3.5.1
7
3.6
3.7
4
Modeling ATIS/VMS with Predictive Information ..............
55
3.5.3
Dissemination of VMS/HAR Information in DynaMIT-P......
56
Framework of the Planning Tool for LWC Case Study ...................
57
3.6.1
Development of Incident Scenarios.................................
57
3.6.2
Evaluation of Diversion Strategies...................................
60
Sum m ary ........................................................................
63
Calibration of DynaMIT-P for the Lower Westchester County
65
4.1
The Lower Westchester County (LWC) Dataset ...........................
65
4.1.1
Network Description ..................................................
65
4.1.2
Available Data Sources ...............................................
67
Calibration Results .............................................................
87
4.2.1
Supply Simulator Parameters ........................................
87
4.2.2
Demand Simulator Parameters .......................................
89
4.2
4.4
5
3.5.2
Sum m ary..........................................................................
98
Case Study
5.1
5.2
99
Scenario 1......................................................................
99
5.1.1
N o Incident .............................................................
99
5.1.2
Incident without VMS .................................................
102
5.1.3
Incident with VMS (Diversion Strategy)............................
106
5.1.4
Comparison of Different Cases: No Incident, Incident without VMS
and Incident with VMS.........................................................
111
Scenario 2 ...........................................................................
114
5.2.1
N o Incident ...............................................................
115
5.2.2
Incident without VMS .................................................
118
5.2.3
Incident with VMS (Diversion Strategy)............................
120
5.2.4
Comparison of Different Cases: No Incident, Incident without VMS
and Incident with VMS.........................................................
5.3
..
Sum m ary ...........................................................................
8
123
126
6
129
Conclusion
6.1
Thesis C ontribution .............................................................
129
6.2
Scope of Future Research ......................................................
130
A
DynaMIT-P Input Files
B
The vms.dat File
C
Analysis of Sensor Data obtained from NYSDOT for LWC Network
9
10
List of Figures
Figure 2-1
The Overall Framework ...................................................
31
Figure 2-2
The Demand Calibration Problem .......................................
32
Figure 2-3
Iterative Demand Calibration Framework ..............................
33
Figure 2-4
Three-stage Supply Calibration ...........................................
35
Figure 3-1
The DynaMIT Overview ...................................................
40
Figure 3-2
System Framework of the Planning Tool using a DTA system .......
48
Figure 3-3
Modeling Day-to-Day Behavior in DynaMIT-P ........................
51
Figure 3-4
Travelers Response to VMS ................................................
54
Figure 3-5
Example to illustrate link-VMS and path-VMS .........................
56
Figure 3-6
Development of Incident Scenario .........................................
59
Figure 3-7
Evaluation of Diversion Strategy using Predictive ATIS Strategy in
D ynaM IT -P ....................................................................
62
Figure 4-1
The Lower Westchester County Network ..................................
66
Figure 4-2
O-D Pairs for LWC Network ................................................
68
Figure 4-3
Location of the Sensors on the LWC Network .............................
71
Figure 4-4
Sensor Count Variation for a Freeway (1-95, North and South) Bound)
S ensor .........................................................................
Figure 4-5
Sensor Count Variation for a Freeway (1-287, East and West Bound)
S ensor ..........................................................................
Figure 4-6
. 75
Sensor Count Variation for a Parkway (Saw Mill River Parkway, North
and South Bound) Sensor .....................................................
Figure 4-7
. . 74
76
Speed-Density-Flow Variation from Acoustic Sensor Data, located
near Exit-5 on 1-287 Eastbound Freeway ..................................
80
Figure 4-8
Hourly Sensor Count Variation ..............................................
81
Figure 4-9
NYMTC Representation of Interchange of 1-87 with Cross County
P ark w ay ..........................................................................
Figure 4-10 Ortho-picture of Interchange of 1-87 with Cross County Parkway
11
82
from G lobeX plorer ..........................................................
Figure 4-11
83
Example of using TransCAD and GlobeXplorer Image to Construct
Detailed Interchange required for the purpose of Traffic Modeling
in D ynaM IT ..................................................................
86
Figure 4-12 Network showing Origin Nodes (1,2,3,4) and Destination Nodes
(5,6,7); Examples of Paths for the OD pair 4-6 .........................
92
Figure 4-13 Travelers' Equilibrium Travel Times (after convergence) between all
OD Pairs "1, 2, 3, 4 - 5,6, 7" ................................................
93
Figure 4-14 Network showing Origin Nodes (1,2,3,4,8) and Destination Nodes
(9,10,11); Examples of Paths for the OD pairs is marked with dark
color........................................................................
. . .. 94
Figure 4-15 Travelers' Equilibrium Travel Times (after convergence) between all
OD Pairs "1, 2, 3, 4, 8 -9, 10, 11" ..........................................
94
Figure 4-16 Estimated vs. Observed Sensors Count from 6:00 AM to 7:00 AM ....
96
Figure 4-17 Estimated vs. Observed Sensors Count from 7:00 AM to 8:00 AM ....
96
Figure 4-18 Estimated vs. Observed Sensors Count from 8:00 AM to 9:00 AM ....
97
Figure 4-19 Estimated vs. Observed Sensors Count from 9:00 AM to 10:00 AM ... 97
Figure 5-1 Network showing Origin Nodes (1,2,3,4) and Destination Nodes
(5,6,7); Examples of Paths for the OD pair 4-6 ..............................
100
Figure 5-2 Travelers' Equilibrium Travel Times (after convergence) between all
OD Pairs "1, 2, 3, 4 - 5, 6, 7" ...................................................
100
Figure 5- 3 Frequency of Experienced Travel Times for all OD Pairs
"1, 2, 3, 4 - 5, 6, 7" ..............................................................
10 1
Figure 5-4 Average Travel Times between OD Pairs "1, 2, 3, 4 - 5, 6, 7" ............. 102
Figure 5-5 Location of the Incident 1 and Examples of Paths for the OD Pair 4-6 ... 103
Figure 5-6 Impact of Incident 1. Travelers' Experienced Travel Times for all OD
P airs ................................................................................
10 4
Figure 5-7 Impact of Incident. Frequency of Experienced Travel Times for all OD
pairs "1, 2, 3, 4 - 5, 6, 7" .....................
.................................
Figure 5-7a Impact of Incident. Average Travel Time for all OD pairs
12
105
"1, 2 , 3, 4 - 5, 6, 7" ............................................................
106
Figure 5-8
Location of VMS Message .....................................................
107
Figure 5-9
Impact of VMS. Travelers Experienced Travel Times for all OD pairs.. 109
Figure 5-10
Impact of VMS. Frequency of Experienced Travel Times for all OD
pairs "1, 2, 3, 4 - 5, 6, 7" ......................................................
Figure 5-11
Impact of VMS. Average Travel Time for all OD pairs
"1, 2, 3, 4 - 5, 6, 7" .............................................................
Figure 5-12
109
110
Network showing Origin Nodes (1,2,3,4,8) and Destination Nodes
(9,10,11); Location of Incident 2 and Examples of Paths for the OD pairs
in dark color........................................................................
Figure 5-13
Travelers' Experienced Travel Times (after convergence) between all
OD Pairs "1, 2, 3, 4, 8 - 9, 10, 11" .............................................
Figure 5-14
116
116
Frequency of Experienced Travel Times for all OD Pairs
"1, 2, 3, 4, 8 - 9, 10, 11" ........................................................
117
Figure 5-15
Average Travel Times between OD Pairs "1, 2, 3, 4,8 - 9, 10, 11" .
117
Figure 5-16
Impact of Incident. Travelers' Experienced Travel Times for all OD
P airs ................................................................................
1 18
Figure 5-17 Impact of Incident. Frequency of Experienced Travel Times for all OD
pairs "1, 2, 3, 4, 8 - 9, 10, 11. .................................................
119
Figure 5-18 Impact of Incident. Average Travel Time for all OD pairs
"1, 2, 3, 4, 8 - 9, 10, 11" ........................................................
120
Figure 5-19 Impact of VMS. Travelers Experienced Travel Times for all OD pairs.... 121
Figure 5-20 Impact of VMS. Frequency of Experienced Travel Times for all OD
pairs "1, 2, 3, 4, 8 - 9, 10, 11. .................................................
122
Figure 5-21 Impact of VMS. Average Travel Time for all OD pairs
"1, 2, 3, 4, 8 - 9, 10,11" .......................................................
13
123
14
List of Tables
Table 4-1
Total (top) and Usable (bottom) Sensors Received from NYSDOT ....
70
Table 4-2
Speed Density Relationship Parameters per Road Type ..................
88
Table 4-3
Final RMS and RMSN Error for Each Time Interval ....................
98
Table 5-1
Comparison based on Aggregate Statistics .................................
111
Table 5-2
Comparison based on the Frequency of Trip Travel Times. For All
O D Pairs "1, 2, 3, 4 - 5, 6, 7" .................................................
Table 5-3
Comparison based on Departure Time Interval. For All OD
Pairs "1, 2, 3, 4 - 5, 6, 7" ....................
................................
Table 5-4
Comparison based on Aggregate Statistics .................................
Table 5-5
Comparison based on the Frequency of Trip Travel Times. For All
OD Pairs "1, 2, 3, 4, 8 - 9, 10, 11" ...........................................
Table 5-6
112
113
124
124
Comparison based on Departure Time Interval. For All OD
Pairs "1, 2, 3, 4, 8 - 9, 10, 11" ...............................................
15
125
16
Chapter 1
Introduction
Transportation systems, especially those involving vehicular traffic, have been subjected to
considerable increases in complexity and congestion during the past two decades. These
conditions often produce traffic situations that jeopardize the safety of the people involved.
Developed countries economies depend heavily on transportation systems and this dependence is
likely to increase as "just-in-time" supply chain management is becoming more and more
popular, that requires reducing the space and investment requirement for storage and rely on fast
and reliable transportation network. Some transportation networks are more vulnerable to
incidents ranging from natural disasters (like earthquakes, floods, storms, landslides and
subsidence) or man-made disasters (like terrorist attacks, bridge or tunnel collapses and major
accidents) at one end, to a variety of events (like congestion, road maintenance, badly-parked
vehicles, and minor collisions). It is difficult to control the scale, frequency and predictability of
these events but it is quite possible to design and manage the transportation system in such way
that disruption caused by such events could be minimized.
It is thus needed to design a Transportation Management System that can handle a range
of traffic conditions, emerging from the incidents mentioned above. In that direction, recent
years have seen a tremendous growth in the field of Intelligent Transportation Systems (ITS) in
parallel with the advancement in Information Technology (IT). Integrating ITS technologies with
traffic surveillance promises to reduce the cost incurred by non-recurrent traffic congestion by
improving transportation agencies' ability to predict and clear such incidents. Advance Traffic
Management Systems (ATMS) and Advance Traffic Information Systems (ATIS) are examples
of ITS sub-systems, which are increasingly becoming prevalent in the Traffic Management
Center (TMC) of a city in order to address various traffic management issues. These tools
17
influence various factors associated with the transportation system like transportation demand (in
reaction to information disseminated to drivers), network capacity (by using speed limit
messages, Variable Message Signs (VMS)) and many other similar components. In order to
make these sub-systems efficient it is very important to estimate and predict traffic condition as
accurately as possible taking into account the impact of provided information to travelers, and
thus provide better guidance to travelers.
Dynamic Traffic Assignments (DTA) systems ([36], [37], [38], [50]) are being developed
to address these issues within Traffic Estimation and Prediction Systems (TrEPS). DynaMIT'
([37], [38]) is a DTA software system that provides travel guidance based on real-time traffic
predictions and drivers reactions to provided guidance. It has been designed to operate in a
Traffic Management Center (TMC) and to support ATIS and ATMS operations. DynaMIT-P is a
planning tool that operates in an offline mode, in which traffic managers investigate and prepare
standard response strategies to different traffic and incident scenarios.
1.1 Background
Traffic congestion in developed and developing countries is putting a lot of pressure on the
economy of a country. Some recent surveys show that traffic congestion costs £15 billion per
year to UK economy [26] and $72 billion per year to US economy [49]. These numbers itself
reinforce the importance of good transportation networks that have minimal traffic congestion
and efficient vehicular movement. A good transportation network not only helps in reducing
traffic congestion but also plays a vital role in the development of a city, state and nation.
Construction of new roads for solving the problem of traffic congestion is not only expensive
and damaging to the environment but it also only provides a temporary solution to the existing
problem. Construction of new roads will attract more traffic, which will nullify the projected
benefits from its construction, making it a cyclic problem. Moreover, it is not always possible to
build new roads due to many political, social and geographical constraints.
Dynamic Network Assignment for the Management of Information to Travelers
18
Above discussion leads us towards the need for better management of the existing road
space with the help of emerging technological tools. Intelligent Transportation Systems (ITS) is
one of the alternative approaches, which can be used to relieve traffic congestion by efficiently
using the available road capacity. Adaptive traffic controls to minimize delays, Ramp Metering
for smooth motorways traffic, Variable Message Signs (VMS) for incident management, speed
control and parking guidance, Dynamic Route Guidance systems for guiding equipped vehicles
to choose fastest path and to respond to incidents in real time, Electronic Toll Collection (ETC)
for smooth revenue collection, are some of the examples which are already in practice to tackle
different traffic scenarios.
A detailed analysis, including technical feasibility analysis and cost benefit analysis
(environment impacts, safety and efficiency), is needed before implementing any one of the
above traffic control alternatives. Most of the public or private agencies are also interested in
evaluating different alternatives. Evaluation of different alternatives to quantify various
performance factors could be quite challenging. One way would be to actually implement the
alternative and then compare the data before and after the introduction of the new system. But
this approach is not desirable because of the following two reasons:
" Implementation of a new system requires a large amount of investment. It might
have severe political and social consequences if the new system is not successful.
*
Many new systems are expected to have modest benefits, e.g. a new urban traffic
control system may reduce travel times by less than 10%. Travel times however vary
a lot from day to day anyway, so it can be hard to determine whether any measured
changes are due to the new system or simply by chance. It is difficult to assess the
benefits of a new system in place because many times it is hard to see if the
measured changes are due to the new system or due to inherent variability in the
network itself.
A more reliable and promising approach will be to use a traffic model to evaluate the
system, in which the traffic modeler has complete control over the different components of the
model. In this way pre and post conditions of a new system can be computed more reliably and
with greater confidence in the results. Traditional traffic models, having simple speed-density-
19
flow relationships, are not able to assess the effectiveness of a new system which often requires
more complicated models. Micro-simulation models, which capture the individual driver
behavior using various car following, lane changing and gap acceptance rules, are becoming
more and more prevalent for the analysis of an ITS system. These models take into account the
response of individual drivers to a change in the network conditions and hence can be used as a
proxy of the real world. Wide range of measures of effectiveness can be obtained as an output of
the simulation in order to provide realistic results, prior to implementation of new system in the
real world. MITSIM [30], PARAMICS [45], VISSIM [44] and CORSIM [21] are some of the
examples of micro-simulation software system available for such kind of analysis.
One of the main problems with micro simulators is that they require a large amount of
computational time for simulating a large transportation network and hence are not well suited
for real-time use for such a network. This necessitates the development of a system that can be
used for the operation of a transportation management system with precision enough for ITS
purposes and is also real time. In this regard, development of a Dynamic Traffic Assignment
(DTA) model has been a subject of major research for transportation experts for more than thirty
years. With the ever-increasing deployment of ITS systems, it has become crucial to come up
with a DTA system which can be used for dynamic traffic management purposes, like reducing
delay on major highways, improving safety, efficiency and capacity of transportation network,
and improving wide-area emergency responses through information sharing and coordination.
With these goals in mind FHWA R&D started a DTA research project in 1994 to address
complex traffic control and management issues in the information-based, dynamic ITS
environment
(http://www.dynamictrafficassignment.org).
DynaMIT,
developed
at
MIT
Intelligent Transportation Systems Program, is a result of this project. It has features like realtime estimation and prediction of network conditions taking into account the drivers response to
disseminated information, traffic information and route guidance to drivers for optimal decisions.
Since DynaMIT has reliable traffic estimation and prediction capabilities, it is well suited
to assist transportation planners and policy makers while evaluating different diversion strategies
in case of special events and emergency situations. Use of DynaMIT during such policymaking
is justified by its unbiased and consistency features, which ensure that DynaMIT's prediction of
20
expected network state matches what drivers would experience on the network. DynaMIT-P,
which is a planning version and meant to assist in offline planning decisions, focuses on shortterm and within-day travel behavior, assuming that the long-term decisions are given.
1.2 Motivation and Thesis Objectives
NYSDOT (New York State Department of Transportation), in association with many public and
private sector companies, is in the process of deploying state-of-the-art ITS installation in the
Lower Hudson Valley (LHV) road network. Existing traffic surveillance and monitoring devices
such as loop detectors and CCTV will be substantially augmented by the deployment of 24
TRANSMIT detectors on the LHV highways and bridges, many of which have high fraction of
EZ-PASS users. With proper processing, the data should make it possible to obtain reliable realtime estimates of traffic flows and travel times by link, path and origin-destination (O-D) pair
over much of the LHV network.
This information can form the basis for providing very effective travel guidance to
network users. Lower Westchester County (LWC) ITS subsystem has considered installing VMS
and Highway Advisory Radio (HAR) transmitter at key network locations. They are intended for
routine guidance use as well as to assist with MPT activities. The combination of NYSDOT's
ITS infrastructure with DynaMIT's guidance generation capabilities would lead to a very
powerful system for improving traffic conditions in the LWC area, and is the motivation for this
thesis.
In a nutshell, the purpose of the proposed thesis is to apply DynaMIT's DTA capabilities
to the LWC ITS subsystem, to generate guidance for the LWC's ITS equipment in an offline
mode, and to use calibrated system to perform illustrative analyses of incident response
strategies. The efficiency of any traveler information and guidance system depends largely on the
accuracy with which the system can estimate and predict network state and path travel times. In
order to ensure that the system reacts and behaves in a realistic manner when used in TMC, it is
critical to calibrate its model components against the field data observed from the network.
Based on the above discussion, the main objectives of the proposed thesis are:
21
i) to calibrate DynaMIT so that it can estimate and predict traffic conditions in the LWC
network with precision sufficient for ITS purposes.
ii) to illustrate the use of calibrated DynaMIT system by applying it to investigate a limited
number of incident scenarios and response strategies in the LWC test bed.
1.3 Thesis Outline
This thesis is organized as follows. Chapter 2 covers the literature available related to calibration
of different components within a DTA system. A calibration methodology for DTA system is
presented with particular emphasis on DynaMIT. It also covers a literature review on modeling
of information dissemination devices. Chapter 3 presents a general framework for the
development of incident scenarios and diversion strategies. It briefly covers components of
DynaMIT-P, a planning version of DTA system, and then explains how it is been used for
developing incident scenarios and corresponding response strategies. Chapter 4 implements the
developed framework to generate several case studies for LWC network. It covers various issues
with available data and assumptions made due to available data quality, with respect to the need
of DTA system components. It presents different calibrated parameters of DynaMIT system for
the LWC network. It finally describes the impact of different incidents and respective diversion
strategies. Chapter 5 summarizes the research work and presents the direction for further work
needed in this area.
22
Chapter 2
Literature Review: Calibration Methodology and
Modeling of Information Dissemination
Traffic assignment models are used to simulate traffic flows on a transportation network, and
various other variables like travel times, pollutant emission, congestion etc. Traditional planning
methods like four-step process mainly consist of static and deterministic models. Currently
agencies use static assignment models for many planning objectives, like improvement of traffic
movement, traffic maintenance, air quality impact etc. However, these traditional tools are not
able to capture the effect of traffic control, information dissemination, traffic dynamics and
driver's response to any information, which make them unsuitable for certain applications,
especially while evaluating evolving and existing ITS technologies. These drawbacks of
traditional planning and assignment models are overcome by Dynamic Traffic Assignment
(DTA) systems, which have evolved substantially in the last two decades and have reached a
sufficient level of maturity to be used in many planning and operational applications. DTA
systems capture traffic dynamics in realistic manner by modeling time dependent OriginDemand (OD) demand, queues and spillbacks. Also, planning version of DTA system can be
used in place of existing planning tools replacing static assignment models.
Despite the fact that uncertainty due to congestion, incidents or any special event should
be accounted in good planning, it has rarely been used in the analysis and forecasting tools which
provide quantitative design decision support for a transportation network. With the evolution of
ITS and DTA systems, it is now possible to design such ATMS/ATIS strategies that will provide
high level of network performance by providing guidance through VMS/HAR, in general
measured in terms of travel times.
23
The Following sections will cover a detailed literature review of DTA systems, its
components and calibration, and need for the development of a planning tool for short-term
planning applications using a DTA system.
2.1 DTA systems
There are mainly two types of DTA systems: (i) Analytical, and (ii) Simulation-based. Analytical
DTA models frame the problem as an optimization and use mathematical optimization
techniques to solve this problem. There are two main disadvantages of analytical DTA system:
(i) these models become very complex and intractable for large and realistic size networks, and
(ii) these models can not capture the realities of network (like street, segment characteristics) due
to the assumptions made to simplify the problem. Simulation-based DTA systems overcome
these disadvantages but they also have several drawbacks like failure to reach convergence.
While more research is needed towards further enhancements and applications of simulationbased DTA systems, they have several advantages:
(a) Simulation-based DTA systems capture some traffic issues which are difficult to capture
in analytical approach, like complex vehicle interactions, traffic signals, realistic network
representation, congestion, queues, spill-backs and congestion dissipation.
(b) Importance of simulation-based DTA systems increases in the context of modeling
various ATMS, ATIS and ITS strategies.
(c) They can capture the time dependent interaction between different components of a DTA
system (like supply and demand). Various travel choices, like mode, departure time and
path, can be modeled easily by simulation-based DTA systems.
2.1.1 Simulation-based DTA systems
Examples of currently available simulation-based DTA systems for planning and decisions
support are DynaMIT, DYNASMART and CONTRAM.
24
DynaMIT, developed at MIT's Intelligent Transportation Systems Program, is the state-ofthe-art traffic estimation and prediction software system, which is designed to reside in a Traffic
Management Center (TMC) for various real-time and planning applications, and to support ATIS
and ATMS strategies. DynaMIT incorporates unbiasedness and consistency to achieve reliable
prediction and credible guidance. Unbiasedness ensures that the information provided to
travelers is based on the best available knowledge of current and anticipated network conditions.
Consistency guarantees that DynaMIT's prediction of expected network conditions match what
drivers would experience on the network. DynaMIT has the ability to trade off level of detail and
computational practicability, without compromising the integrity of its output. Its important
features and functionality include:
(a) Demand simulation using a micro-simulator, which generates individual travelers and
simulates their choices regarding whether to travel or not, departure time, mode, and
route (pre-trip and en-route) in response to information provided by the ATIS.
(b) Supply simulation using a mesoscopic traffic simulator that explicitly models the traffic
dynamics of development and dissipation of queues, spillbacks, and congestion in
general.
(c) Individually simulates each trip, to generate detailed vehicle trajectories, which is very
useful to various ATIS requirements.
(d) Uses a rolling horizon to achieve efficient and accurate real-time estimations and
predictions. It employs this predicted information to generate descriptive and prescriptive
information that avoids overreaction to any incident.
(e) DynaMIT-P, planning version, is designed to assist the evaluations of infrastructural,
operational, or informational changes to local and regional transportation networks, by
efficiently modeling day-today and within-day patterns of traffic flows, travel demand
and network conditions.
DYNASMART, developed at University of Texas at Austin, is an assignment-simulationbased framework that combines advance network algorithms and models trip-maker behavior in
response to information, using real-time information, such as loop detectors, roadside sensors,
and vehicle probes. Consistency checking and updating is an important function incorporated in
25
DYNASMART to ensure consistency of the simulation-assignment model results with actual
observations, and to update the estimated state of the system accordingly. Key Features of
DYNASMART include:
(a) A simulation-based dynamic traffic assignment system with micro-simulation of
individual user decisions in response to information and a mesoscopic traffic flow
simulation approach.
(b) Recognizes multiple user classes in terms of (1) operational performance (e.g., trucks,
buses, passenger cars), (2) information availability and type, and (3) user behavior rule
and response to information.
(c) DYNASMART-P, planning version, can model several features, like ability to load
individual trips and trip chains with several intervening stops of associated durations,
evaluation of HOV[HOT pricing schemes, VMS and incident management.
CONTRAM (CONtinuous TRaffic Assignment Model), developed by Transportation
Research Laboratory (TRL) and Mott Macdonald, UK, has the capability to model variety of
situations from congested urban networks to regional inter-urban areas. It can model unexpected
events such as incidents that reduce network capacity and the effects of driver information
systems. Fuel consumption and pollutants emission can also be modeled by CONTRAM. By
interfacing with TUBA ( Transport User Benefit Appraisal) it can do the cost benefit analysis of
new infrastructure changes and traffic control measure. Special vehicle lanes, like HOV/HOT,
modeling or banning turning movements in certain situations can be handled by it. Dynamic
modeling is done by CONTRAM by dividing the day into time slices to model peak and off-peak
traffic conditions. Vehicles are assigned to minimum cost paths by taking into account the
vehicle interactions and delays caused by these interactions. Some of the key modeling features
are: dynamic matrix estimation, an unlimited number of time slices, up to 32 user definable
vehicle classes, left or right hand drive. One of the major limitations of CONTRAM is that it
cannot explicitly capture the individual driver behavior and decisions, which are very crucial for
evaluating different ATIS/ATMS strategies.
26
2.1.2 Model Components in DTA system
A general framework of the DTA system consists of three sub-models:
(1) Demand Model: simulates the effects of the transportation system state on the
behavior of users.
(2) Supply Model: simulates the effects of user behavior on the transportation system.
It captures the traffic dynamics (like congestion, queues, spill-backs, travel times,
flows, speed densities) at all points on the network, and hence used to indicate
network performance.
(3) Demand-Supply Interaction Model: simulates the interaction between the above
two models (day-to-day and within-day models)
(1) The Demand Model
Demand model in DTA systems captures the driver behavior with respect to pre-trip or en-route
information on departure time, mode and route choice, as well as day-to-day behavior modeling
of these choices. These choices are represented as a function of different kinds of attributes like
travel time, cost, comfort and other socio-economic factors like age, sex, income level etc. A key
component of demand model is OD estimation, which involves adjustment in demand in order to
match traffic sensor counts as closely as possible. A general framework for dynamic demand
simulation is presented by Ben-Akiva et al. [11]. A detailed literature review about the different
modules of the demand model is done by Sundaram, S. [48]. A brief discussion of these
components and their importance with regard to DTA system is presented below:
(a) The Route Choice Model
The route choice model captures the important aspect of dynamic traffic assignments. It
contains choice set of all the feasible routes for each driver. Most of the route choice
models use Multinomial Logit Model (MNL) for calculating the choice probability for all
feasible routes of each user. It then assigns the driver to the path with maximum
probability and utility. The Independence from Irrelevant Alternatives (IIA) limitation of
27
MNL model is usually resolved by C-logit route choice model (Cascetta et al. [17]),
which account for amount of overlapping among paths and hence give more realistic
routes flow fractions. Path-size (PS) logit model by Ramming [46] uses the correction
term for overlapping that takes into account behavior theory and uses discrete choice
theory (Ben-Akiva and Lerman, [8]).
Route choice can be modeled at Pre-trip and En-route level based on the ATIS
information. One of the major differences between these two kinds of models is the fact
that shift in departure time is no longer available for en-route choice model because
driver has already entered the network. Modeling of route choice in response to
information, available through VMS or through some dedicated system, is usually done
in two ways. One of them is to take into account the factors into discrete choice
modeling, which are considered to influence en-route decisions. Other way is to model
these choices by bounded rationality concept, which states that driver's decision depends
on the satisfaction level (Simon [47], Mahmassani and Chang [35], Mahmassani et al.
[36]). For example, driver will change the route only if the total saving in travel time is
above some threshold value. An approach using fuzzy theory to model route-choice in the
presence of information is also presented by Lotan and Koutsopoulos [33].
(b) The Departure Time Model
Departure time choice of driver is sensitive to information provided regarding incident,
congestion, and some personal attributes like value-of-time and preferred time of arrival.
Generally used models for departure time choice are Logit model, Generalized Extreme
Value (GEV) model and Logit kernel probit etc. Similar to route choice model, these
models suffer from IIA property especially when departure time choices are closer, in
which case error term may be correlated. To overcome this problem, departure time
switching models are employed where choices are either habitual departure time or, some
earlier or later time slot.
(c) The Mode Choice Model
Most of the mode choice models are based on logit or nested logit models, which use
random utility principle. To choose a mode like car, transit, bus or any other mode
28
depends on various explanatory variables like travel time, travel cost, comfort, number of
changes, income level, age and many other socio-economic attributes.
(d) The O-D Estimation Module
The O-D estimation module estimates the OD flows by varying time interval so that
simulated sensor counts and field counts match within desired limits. Generally used
algorithms for OD estimation are either simultaneous or sequential. Some of these
approaches are the Kalman Filter approach and Generalized Least Square (GLS)
approach. Balakrishna [6], Ashok [4] and Brandriss [13] have described these two
approaches in detail. Some of the very good and comprehensive work in this area is been
presented by Ashok and Ben-Akiva [5], Cascetta [14] and Cascetta et al. [16].
(2) The Supply Model
The supply model of the DTA system captures the traffic dynamics of the network by simulating
the movement of drivers. It does not capture the detailed drivers interactions, like lane changing,
car following models in micro-simulation, but primarily captures queuing behavior of the traffic.
It provides network performances in terms of flows, speeds, densities, travel times, and build-up
and dissipation of lane-specific queues and spillbacks over all the points on the network.
Most of supply simulators are mesoscopic in nature based on two types of models: 1. a
deterministic queuing model and, 2. a moving model. The queuing model simulates the
spillbacks by using input and output capacity of the segment. Insufficient capacity or space on
the current or downstream segments will cause vehicles to start queuing up. The moving model
is based on macroscopic speed-density relationship. Further, capacities are also used to model
incidents and signalized intersections. Detail literature on the supply simulation can be found in
Ben-Akiva et al. [9] and Kunde [31].
(3) The Demand-supply Interaction Model
29
The Demand-supply interaction model captures the day-to-day learning experience of the users.
This behavioral principle states that choices of the users depend on the expected pre-trip travel
times. Expected value of the travel time is a result of experience, memory and learning. Cascetta
and Canterella [15] have described day-to-day stochastic assignment models, which predict the
driver's expected travel time based on his experienced and expected travel times in the past days.
An equilibrium model, where travel choices are based on the weighted average of experienced
travel times during the previous days, is proposed by Horowitz [25]. Ben-Akiva et al. [10] and
Jha et al. [28] have presented models which incorporate information provided to drivers into the
day-to-day learning models. Mahmassani and Chang [34] have proposed model which takes into
account the late or early arrival of driver on the previous days.
2.2 Calibration Methodology for DTA systems
The efficiency of any traveler information and guidance system depends largely on the accuracy
with which the system can estimate and predict network state and path travel times. This leads to
level of detail and accuracy of the two simulation tools used by DTA systems: the demand
simulator and the supply simulator. In order to ensure that the systems react and behave in a
realistic manner when used in TMC, it is critical to calibrate its model components against the
field data observed from the network.
2.2.1 Overall Calibration Approach
The generalized approach for the calibration of a DTA system can be summarized as shown in
Figure 2-1. The process begins with the initial values of calibration parameters based on
available data or on some reasonable assumptions. These initial parameters are fed into DTA
system to generate model outputs. An objective function will then be computed based on model
outputs and observed field data. The next set of parameters values will be determined by an
optimization step, which will again be used in DTA system as new inputs. After the
convergence, model outputs will be used for further analysis.
30
The Optimization problem mentioned in the above framework is usually very large and
complex, and might therefore be solved sequentially. The calibration process can be seen as an
iterative sequence of demand and supply model calibrations within the DTA system, with their
interaction resulting into final model outputs in each iteration. The Following sections will
present the approaches to calibrate the demand and supply models within DTA system.
In itia I
V a Iu es
D TA System
M
Updated
odeI
Param eter
Values
utputs
0Observed
D ata
01 0bjective
F unctio n
Op tim
iz atio
N
n
o
Con v erg enc
C
e
Analysis
Figure 2-1: The Overall Framework
2.2.2 Demand Calibration
Demand Calibration Elements. Demand calibration methodology will integrate the route
choice, departure time choice and O-D estimation modules into a single united frame. Input and
output requirements of the demand calibration are shown in Figure 2-2. This process uses
31
available estimates of O-D flows and several days of sensor counts and occupancies. The route
choice parameters and O-D flows estimation models inputs are the outputs of this process.
a priori O -D
Flows (From
Surveys/M odels)
'N' days of
sensor data
Route Choice Parameters
Time-Varying Historical
O-D Flows
Error Covariance Matrices
Autoregressive Factors
Figure 2-2: The Demand Calibration Problem
In summary, the demand simulator's parameters which need to be calibrated can be
grouped as below:
Route Choice Parameters
i)
Parameters in the path choice set generation algorithm
ii)
Parameters in the path utility specification
iii)
Path-size exponent
O-D Estimation (and Prediction) Parameters:
i)
The historical database of O-D flows, XH
ii)
The variance-covariance matrix Vh associated with indirect measurement errors
iii)
The variance-covariance matrix Wh associated with direct measurements errors
iv)
The matrices
ff
of autoregressive factors
Balakrishna [6] has described that demand calibration problem becomes a fixed point
problem between the route choice fractions, the assignments fractions and the O-D flows, and
can be solved by iterative approach. The paper has described in detail an iterative approach
(Figure 2-3) to demand calibration where route choice model has been used as the starting point
in the iterative framework.
32
Start
Path Travel times
Route Choice
Model
Update Route
Choice Model
Parameters
i
4
No
O-D Flows
Dynamic Network
Loading (DNL)
Model
No-
Assignment
Estimation
P O-D Module
Matrix
Convergence
Route Choice
Yes
Parameters
Convergence
Yes
Stop
Counts
Figure 2-3: Iterative Demand Calibration Framework.
2.2.3 Supply Calibration
Supply Calibration Elements. The supply simulator in DTA system (in particular in DynaMIT)
is a mesoscopic traffic simulator that simulates vehicular movements and provides matrices of
network performances such as the temporal and spatial flows, speeds, queue lengths, and travel
times at all points on the network. The supply simulator obtains the network description and the
list of packets (vehicles) to be moved on the network through its interfaces with the network
topology component and the list of packets component respectively. It then simulates the
33
movement of vehicles on the network for the given supply simulation time interval. Calibrated
speed-density relationship parameters and lane group parameters (capacities) are the outputs of
supply calibration module.
The list below shows the summary of parameters which are key to the calibration of supply
simulator:
i)
Segment-specific speed-density parameters u, umin, ko, kjarn, a and
ii)
Lane group capacities on freeway and arterial segments, and
iii)
Lane group capacities at the signalized intersections.
Q,
Kunde [31] has shown in detail the three-stage calibration methodology (Figure 2-4) for
supply simulator parameters. The approach divides network into three different levels, (i)
Disaggregate level (ii) Sub-network Level, and (iii) Entire Network level. Calibration of
parameters is then done from microscopic level (disaggregate) to macroscopic level (entire
network).
As mentioned before, since calibration of demand and supply parameters is a very large and
complex problem calibration of these parameters is usually done iteratively between demand and
supply models. While demand simulator needs specific supply model parameters (like link travel
time) as an input, supply simulator moves the vehicles on the network assuming driver choices
are known. Hence, calibration process can be seen as an iterative sequence of demand and supply
model calibration, with their interaction resulting in the final model outputs in each iteration.
34
Traff ic Sensor Data
Calibration of Individual elements
1. Speed-density relationships of
segments
2. Capacities of segments (accounting
for intersections)
True O-D flows,
Route Choice
Subnetwork
Calibration
Initial
Paramter Values
System level calibration
Supply
I
Demand
Figure 2-4: Three-stage Supply Calibration
2.3 DTA System for Short-Term Planning Applications
The main limitations of the static planning tools (like EMME2, TransCAD, TRANSIMS,
TRANPLAN, etc.) for short-term planning applications, especially in the presence of ITS, are
their inability to capture traffic dynamics,
dynamic demand-supply
interactions and
characteristics of ATIS/ATMS. Furthermore, the network performance is due to the individual
decision of the travelers and it is necessary to use disaggregate models to capture traveler
behavior. Existing planning tools do not capture such a rich and explicit modeling of traveler
35
behavior. One of the main reasons for the need of a new planning tool is to address ITS in the
planning context. Currently there are very few tools that are useful for analyzing ITS
deployments. IDAS (Intelligent Deployment Analysis Software) is a recent tool developed by
FHWA. However, existing planning tools (including IDAS), which are based on static
techniques, are not very effective in evaluating ITS strategies.
A brief review of some of the studies for evaluating ITS and other strategies, and the
limitations of the techniques used to evaluate them are summarized below. A more detailed
treatment of the literature review on this topic can be found in Sundaram [48].
0
ATIS Strategies:
It has been pretty well established in the literature that provision of information is
effective primarily under non-recurrent congestion conditions.
Al-Deek and Kanafani
[2], modeled the benefits of information in traffic corridors and based on a simulation
study reported that travel time savings are significant and can be in the order of 30
percent in a simple network with two routes. Simulation results in the Santa Monica
Freeway corridor in Los Angeles (Gardes and May, [22]), showed that in the presence of
an accident, information provision can bring about 6.2 percent reduction in travel times.
However, a few authors have questioned the significance of Advanced Traveler
Information Systems (ATIS) benefits (e.g. Arnott et al. [3], Hall [23]). Arnott argues that
ATIS may counter-productively lead travelers to congested alternative routes.
Hall
questions whether the problem of non-recurrent congestion is as significant as claimed.
The general conclusion, however, from the available literature is that ATIS can be
beneficial but this statement has to be verified, either by observing actual ATIS
deployments or by conducting further studies.
By modeling in-vehicle information,
VMS and HAR devices, DynaMIT becomes a valuable tool for such studies.
Clearly, to analyze the impact of information either through in-vehicle guidance,
VMS messages or other means, it is necessary to explicitly model both traveler behavior
and traffic dynamics accurately. Existing simulation tools lack one or more of these
critical features. Specifically regarding in-vehicle information, several authors have tried
to analyze the impact of in-vehicle information with market penetration (e.g.
36
Jayakrishnan et al. [27], Walting and Van Vuren [51]). However, studies with respect to
market penetration effects were based either on surveys or on queuing models and do not
effectively capture traffic dynamics and traveler behavior in response to information.
Impact of VMS messages also typically depends on driver response to these
messages and the resulting traffic conditions.
Response behaviors of drivers in the
presence of a VMS have been addressed by the use of logit models based on SP (Stated
Preference) surveys (e.g. Bonsall and Merall, [12]). The main drawback of these surveys
relates to the well-known problems with SP surveys (e.g. response bias, justification
bias). An example of this issue is the study conducted by Chatterjee et al. [18]. As
reported in that study, only one-fifth of the drivers indicated by SP surveys actually
diverted in one area, while SP results were consistent with observed diversion rates in
another location.
Under VMS information, studies based on SP surveys by Wardman et al. [52],
Bonsall and Merall [12] and Peeta et al. [43] indicate that the following parameters are
important factors that govern travelers' route choice decisions: relative journey times,
delay on the current route, age, sex and previous network knowledge. McArthur [40]
used behavioral rules employed in PARAMICS-CM and found that diversion is based on
whether the savings that travelers perceive lie above a threshold, as well as on the
travelers' patience and trust in the system.
9
Demand Management Strategies:
Common example of demand management strategies is the use of HOV[HOT lanes.
Several studies (Dahlgren [19], Johnston and Ceerla [29]) have tried to assess the impact
of HOV/HOT lane strategies. However, these studies do not explicitly consider traffic
dynamics and it is very difficult to analyze for example a strategy such as converting an
already existing general-purpose lane into a HOV/HOT lane. Thus there is need to
develop tools that would capture not only rich traveler choice models to estimate the
mode shift associated with HOV lanes but that would capture traffic dynamics that will
determine the network impacts of such strategies.
37
Based on the preceding discussion, there is indeed a strong motivation to develop planning tools
that will encapsulate traveler behavior, capture traffic dynamics and can be used for a variety of
planning applications (including ITS strategies). The shortcomings of the traditional static
planning tools are overcome by DTA systems, which capture traffic dynamics in a realistic
manner by modeling time varying OD demands, queues and spill-backs. Sundaram [48] has
further developed a framework for short-term planning applications using a DTA system. This
framework is able to model day-to-day evolution of travel demand and network conditions and
capture the within-day dynamics in the case of stochastic events. With current emphasis on
ATIS/ATMS strategies and ITS investments to improve traffic conditions, the developed
planning tool is capable of addressing these issues. This developed tool will be used, in the
current thesis, for evaluating diversion strategy under incident response using VMS/HAR. The
details of this planning tool are given in the next chapter.
2.4 Summary
In this chapter, first we have described different types of DTA systems, its main features and
different kinds of models used within the DTA systems. The importance of calibration while
implementing a DTA system and different calibration approaches for calibrating various supply
and demand parameters were discussed next. Finally, this chapter focused on the modeling of
information dissemination devices, which is very important and critical while supporting a
decision system, especially in case of congestion and incident. It was emphasized that current
DTA systems, like DynaMIT and DYNASMART, take into account the reaction of drivers to the
information provided. Therefore these DTA systems, combined with ATMS/ATIS tools, can be
used for better performance of network, especially in terms of reliability of travel times.
The next chapter presents a framework for the development and evaluation of diversion
strategies
in
case
of
incidents
using
a
38
state-of-the-art
DTA
system
(DynaMIT).
Chapter 3
Development of Incident Scenario and Evaluation
of Diversion Strategy using DynaMIT-P
This chapter presents the framework for developing incident scenarios and diversion strategies
using the DTA system DynaMIT-P, a planning version of DynaMIT. A brief overview of its
features and functionalities were given in Chapter 2. This chapter will provide a deeper
understanding of its main components.
A description of some of the modeling features that
make DynaMIT-P suitable for this kind of planning applications are covered next, especially
day-to-day and within-day driver behavior modeling. Finally, representation of incident
scenarios and representation of VMS/HAR for response strategies in DynaMIT-P have been
described.
3.1 DynaMIT: An Overview
The overall framework of DynaMIT showing its various components is shown in Figure 3-1. For
detailed description of DynaMIT, reader is referred to Balakrishna [4]. It is designed to reside in
a Traffic Management Center (TMC) for supporting various ATIS/ATMS operations. As clear
from the figure, DynaMIT takes both real-time and off-line data for its operation. The off-line
data consists of network description, time-dependent O-D matrices, link travel times and default
parameters for different models. The real-time information consists of surveillance data, traffic
control information and incident characteristics (like location, starting time, duration and
severity).
39
Demand
Simulation
Supply
Simulation
Demand
supply
Simulation
Simulation
Information
Generation
Figure 3-1: The DynaMIT Overview
The state estimation module gives the current estimate of the network in terms of O-D
flows, speeds, densities, queues and link flows, using the inputs (sensor counts etc.) from the
network. The State Estimation module has two main models: (1) The Demand Simulator, and (2)
The Supply Simulator. The Demand Simulator has the capabilities to do real-time O-D
estimation, taking into account the user behavior for route, departure time and mode. It also
models the impact of information provided to user by modeling the reaction of each individual to
the given information, and hence updates the historical O-D matrices. The O-D model then uses
the real-time sensor counts, updated O-D flows, assignment matrices (fraction of OD flows to
link flows) to estimate the current interval O-D flows. The Supply Simulator simulates the traffic
conditions over the network using the estimated O-D flows from the Demand Simulator, updated
capacities (due to incidents etc.), traffic dynamics parameters, traffic control strategies and traffic
information and guidance disseminated (using VMS, HAR or any other dedicated guidance
40
system). Response of users to ATIS is captured through en-route driver behavior model.
Demand-supply interaction is an important component of estimation module. The O-D
estimation model uses assignment matrix that is an output of the Supply Simulator. But
assignment matrix itself depends on route choice decisions made by drivers, which is a part of
the Demand Simulator. Hence, it became necessary to iterate between the Demand Simulator and
the Supply Simulator in order to achieve convergence.
The Prediction-based
Information Generation module uses the predicted traffic
conditions to generate information and guidance to users using various ATIS strategies. An
iterative framework is used to obtain a guidance that should be the best under such conditions.
This important property is called consistency that ensures that there would be no better path that
the driver could have chosen. Each iteration consists of a trial strategy, network state prediction
(both demand and supply prediction) under the strategy and evaluation of the predicted state for
consistency. After the consistency is been achieved, information is disseminated to users in the
network.
3.2 DynaMIT-P: Planning Version of DynaMIT
DynaMIT-P is the planning version of DynaMIT, which has the potential to substantially
improve short-term transportation planning process, especially involving congested network. It is
designed to assist the evaluations of proposed changes to local and regional transportation
networks. Such changes could be infrastructural, operational or informational in nature.
DynaMIT-P is the efficient adaptation of DynaMIT to achieve short-term planning process
objectives.
DynaMIT-P models short-term (day-to-day) and within-day travel decisions. Short-term
travel decisions include departure time, route, mode and destination choices. These short-term
decisions are also affected by long-term decisions. For example, a long-term decision about
residential location, auto ownership or changes in the network will affect the short-term
decisions like departure time, route etc. With their daily travel choices and behavior, individuals
develop habitual travel patterns that they follow regularly. Within-day decisions capture the
41
deviation from the habitual travel pattern, in response to real-time information, weather
conditions, special events or incidents. Since the objective of this thesis is to develop incident
scenarios and diversion strategies for LWC network, DynaMIT-P is well suited for this purpose.
Before we go into the framework for developing incident scenarios and diversion strategies, the
main components of DynaMIT-P are described below.
3.2.1 Main Components of DynaMIT-P
The main components of DynaMIT-P are the demand simulator, the supply simulator and the
learning models using demand-supply interactions.
3.2.1.1 Demand Simulator
The demand simulator in DynaMIT-P is a microscopic simulator, which captures driver behavior
models and OD estimation models. The demand simulator consists of following main models:
(a) O-D Estimation
(b) Demand Disaggregation
(c) Travel Choice Models
(a) O-D Estimation
The current version of DynaMIT-P uses a sequential GLS-based O-D estimation module that can
be expressed by the following equation:
Xh
= argmin[(h-~)w
=gxWh
x
~-h-1
h-1
h
h
h
h
xh
VhYh
p=h-p
where,
Xh=
x
Estimated flows for interval h,
= Target OD flows,
Wh = Error covariance matrix associated with OD flow measurements,
42
hp
p=h-p
h
(3.1)
Yh
= Sensor counts for the time interval h,
ap = Assignment matrix mapping OD flows to link flows from departure interval p to current
interval h,
Vh =
Error covariance matrix associated with the link counts and
p = number of intervals needed for longest trip.
Under the constraint, all the OD flows should be non-negative.
As it is clear from the above equation, OD estimation tries to minimize the deviation of
target OD flows and sensor counts from their estimated values. The key inputs to OD estimation
module are sensor counts, time dependent historical OD flows and assignment matrix. While
sensor counts and OD flows are provided externally, the assignment matrix is generated inside
the supply simulator.
(b) Demand Disaggregation
Before the demand is loaded on the network, it is important to generate the population of drivers
from the OD matrix, by assigning origin, destination, departure time and mode. The origin and
destination is easily obtained from OD matrix, departure time is obtained from corresponding
time interval and mode is assumed to be car by default. A number of socio-economic attributes
(like informed/uninformed driver, value of time) and trip purpose (such as work) are also
assigned to each driver.
(c) Travel Choice Models
Travel choice models within the DynaMIT-P can further be categorized as follows:
(1) Path Choice Set Generation
This is one of the crucial steps towards the route choice model and calibration, because in
congested conditions drivers might be looking for routes other than shortest paths in uncongested
conditions. This step generates a good set of feasible paths between each O-D pair. It consists of
three steps. First step involves the generation of shortest paths between each OD pair. This will
be the most probable path during uncogested conditions. Second step is the link elimination step,
43
which generates alternative paths from shortest paths. This is done by eliminating each link one
by one and by regenerating shortest paths. Generation of alternative paths will also be crucial
during incidents where driver needs to choose alternative path. Third step is the random
perturbation step which perturbs the impedances of the links randomly to capture time varying
link travel times, hence again creating the richer path set. Finally, all the paths are checked for
uniqueness and unreasonably long paths are removed from the path set.
(2) Route Choice
After the path choice set has been generated, assignment of route choice to each driver is done by
Path-Size Logit (PS-Logit) model (Ramming [46]) that can be defined mathematically as
follows:
Vi +'n Psi
(3.2)
+
n
je C,
where,
, (i) = the probability of user n for choosing route i,
Vi = the utility of alternative i,
PS = the size of the path i,
C, = the choice set for user n.
The utility Vi of each path is a function of many explanatory variables, like path travel times,
travel costs and various socio-economic attributes of the users. The default utility function used
in DynaMIT-P is given below:
V
(3.3)
=A$ttsi + (fl2A1)tt Fi
where
ttAi
and ttFi are the arterial and freeway travel time of the total travel time on route i.
and f 2 are the coefficient of arterial travel time and freeway bias, respectively.
The Path-Size PSin is defined as:
44
Q
(3.4)
PS in,= -Ia
aEaj
jE C,
where la is the length of link a, Li is the length of route i, 6 aj is either 1, if link a is part of route j,
or 0 otherwise. The inner summation is over all the routes in the choice set C,, while the outer
summation is over all the links in the path FI. y is known as the path-size exponent.
(3) Departure Time Choice
The version of DynaMIT-P used for this thesis does not have a departure time choice model.
Travelers are assigned a specific departure time based on the uniform distribution over a
departure interval.
(4) Mode Choice
DynaMIT-P supports alternative modes such as HOV vehicles, but the default mode in the
DynaMIT is assumed to be the car.
3.2.1.2 Supply Simulator
The supply simulator in DynaMIT-P is a mesoscopic simulator with the following key features:
*
Movement of the vehicles in the uncongested uninterrupted part of the network is governed
by a macroscopic speed-density relationship given by:
v =Max(
vin, v
1-
[imax(, k -kmn)
)
"
(3.5)
kjamn
where v is the speed of the vehicle (in mile per hour), v,a is the free-flow speed on the
segment, k is the current segment density (in vehicles/mile/lane), kmi, is the minimum density
after which free-flow conditions begin to break down, kjam is the jam density, and cx and $ are
the segment specific coefficients. vi'in is the minimum speed of the vehicle in the moving part
which is determined by network and vehicle attributes.
45
*
The movement of vehicles from one segment to another is controlled by input and output
capacities of the segments. Queuing will start building up if there is not required input/output
capacity to process the vehicles. It employs deterministic queuing model to reflect the
bottlenecks. Incidents and intersections controls are achieved through capacity controls.
Further, the lanes within each segment are grouped into different lane-groups to capture the
effect of turning-movement-specific capacities at diversion and merge points and at
intersections.
*
The simulation of traffic in the supply simulator is done in two phases: the update phase and
the advance phase. The update phase computes the most time-consuming traffic performance
measures like speeds, densities and flows. The advance phase moves the vehicles to their
new positions. The durations of the update phase and the advance phase depend on the trade
off between the accuracy of the result and the computational time.
3.2.1.3 The Learning Model
The learning model in the DynaMIT-P updates the drivers' perceptions of travel times based on
the past experiences, according to the model given below:
k
=
(3.6)
ATk-' + (1 -A)f-lI
where Tk' is the time-dependent travel time along path k on day t, T is time-dependent travel
time experienced along path k on day t, and X captures the learning rate. The value of X lies
between 0 and 1, and is affected by the ATIS strategies.
3.3 DynaMIT-P for Short-term Planning Applications
In order to efficiently simulate the impact of an incident or a diversion strategy, a DTA system
should have functionalities to model realistic day-to-day and within-day driver behavior. Day-today behavior tries to establish the network conditions resulting from traveler's habitual
decisions. It generates the network conditions that would be representative of a normal day
46
without any accidents, special events etc. The within-day dynamics is the next stage of a
planning tool that captures the network state in case of any unusual events. It disturbs the day-today equilibrium conditions and traffic control strategies. DynaMIT-P efficiently achieves dayto-day and within-day behavior modeling of a planning tool as shown in Figure 3.2. Hence, it is
very suitable for the development of incident scenarios and diversion strategies. The outputs
from this kind of planning tool are various measures of performance of the network such as
travel delays, travel costs, fuel consumptions etc. DynaMIT-P's open system of demand models,
detailed representation of network dynamics through supply simulator, learning models, and
flexible structure make it a useful tool for many such planning applications.
47
Historical OD Matrix
Day.-to-Day Behavior
Sensor
Counts
Demand Simulator
Supply Simulator
- Destination
- Departure Time
- Mode / Route
- Control Strategies
- Segment Capacities
- Frequency / Purpose
Learning Models
-
Travel Attributes +No
Information
Equilibrium ?
Yes
Habitual Travel Patterns and Updated OD Flows
Stochastic Factors
4
Within-Day Behavior
- Weather
- Incidents
-Special
Events
ATMS
Supply Simulator
Demand Simulator
- Route Change
- Segment
- Mode Change
Capacities Changes
- Departure Time Change
- Add / Cancel Trip
ATIS VMS
-Information
Expected Resources
- Level of Service
- Safety
- Energy Consumptions
Figure 3.2: System Framework of the Planning Tool using a DTA system
48
3.3.1 Modeling Day-to-Day Behavior in DynaMIT-P
We will now briefly review how DynaMIT-P models the day-to-day behavior. Sundaram [48]
explained in detail the modeling of day-to-day and within-day behavior using DynaMIT-P.
Figure 3.3 gives an overview of the implementation of day-to-day behavior modeling in
DynaMIT-P. As shown in the figure, the inputs required for the day-to-day behavior modeling in
DynaMIT-P are: time dependent historical OD matrices, time dependent historical link travel
times, field sensor counts, socio-economic characteristics. The outputs from DynaMIT-P after
establishing day-to-day behavior are: equilibrium travel times and planning OD matrices. The
convergence criterion used in DynaMIT-P is given by:
F
TI
$
,(SC' -I FC)'
FN<)
E
(3.7)
where SC' are the simulated counts reported by DynaMIT-P for sensor s in time interval t, FCst
are the corresponding field sensor counts, N, is the number of sensor reporting counts, N, is the
number of time intervals and E is the threshold parameter defined by the user.
Two of the most important modules used by this framework are "Equilibrium Travel
Times Computation" and "OD Estimation". Readers are referred to Sundaram [48] for more
details on these modules. The equilibrium travel times are established by an iterative procedure,
until the traveler's expected travel times (i.e. the previous estimate of the equilibrium travel
times) match their experienced travel times (generated by the supply simulator). The
convergence criterion for "Equilibrium Travel Times Computation" tries to minimize the
difference between the previous equilibrium travel times and the latest travel times obtained
from the supply simulator, and is given by following equation:
I
T(EQ't
1=
- EQ: 1
)2
< E(3.8)
where EQ'' is the equilibrium travel time in the current iteration i, on link 1, for a vehicle that
enters the link in time interval t, EQ'' is the corresponding travel time in the previous iteration
49
i-1, N, is the number of links in the network, N, is the number of time intervals with which the
travel times are represented and E is the threshold parameter defined by the user.
The "OD Estimation" module updates the planning OD to reflect the field sensor counts.
As discussed in the chapter 2, problem of OD estimation is a fixed-point problem. The solution
of OD estimation is obtained by an iterative process between the OD estimation algorithm and
the supply simulator, linked by an assignment matrix. The assignment matrix consists of the
fraction of each OD pair that has been counted on each sensor by the time period. The
convergence criteria used for OD estimation is given by:
"F,"' ' - F i'
N
od
)
(3.9)
<
where Fj,dp is the OD flow in the current iteration i, from origin o to destination d in time
interval p, F,id,p is the corresponding OD flow in the previous iteration i-1,
number of OD pairs and E is the threshold parameter defined by the user.
50
Nod
is the total
START
it
Historia
Initial OD Matrix=
OD Matrix
:I_0
OD
Historical
INPUT
Historical
Travel-Times
'
Equilibrium Travel
Times =Historical
Travel Times
STEP 1: Compute Equilibrium TravelTimes given the Initial OD Matrix
iel
Sensor
uns
CFCounts
STEP 2: Estimate OD Matrix given the
historical OD as the seed, Equilibrium
Travel-Times and the Field Sensor
Counts.
Convergence?
No
Update 00
Yes
OUTPUTS
Updated historical
OD and Equilibrium
Travel-Times
ISTOPD
Figure 3.3: Modeling Day-to-Day Behavior in DynaMIT-P
51
3.3.2 Modeling Within-Day Behavior in DynaMIT-P
Modeling the within-Day behavior becomes important especially in the case of special events,
accidents or bad weather conditions. From the Figure 3.2, it is clear that within-day modeling
takes equilibrium travel times and estimated planning OD matrix, from the day-to-day analysis,
as inputs. Within-day behavior is obtained by simply doing iterations between the demand
simulator and the supply simulator in response to stochastic events or ATMS/ATIS strategies.
Implementation of within-day dynamics may require changes on both the supply and the demand
sides. Most of the changes on the supply side of DynaMIT-P are captured through segment
capacities and changes on the demand side are reflected through the planning OD matrix or
through the travel choice models.
3.4 Representation of Incident Scenarios in DynaMIT-P
The day-to-day model captures incident scenarios. These scenarios are captured by modifying
the relevant inputs on the supply or the demand side, and evaluating the network performances
under modified inputs. Changes on the supply side are made primarily on the segment capacities.
For example, lane closure due to an accident and its impacts for a temporary period are captured
by reducing the capacity of the corresponding segment for required time interval. Other similar
examples could be the case of work-zone management, severe weather conditions, etc. Any
changes on the demand side are performed by updating the OD matrix. A typical example will be
the increase in demand due to events like sports, concerts, special events etc. Once the inputs
have been modified to suit particular scenario, the scenario is evaluated by disaggregating the
appropriate OD matrix using the equilibrium travel times to evaluate users' choices. The supply
simulator then yields the network conditions. Several iterations of the demand and supply may
be required to account for the users' adjustment made in response to network conditions. The
impact of the scenarios is analyzed by comparing the normal day network state with the obtained
network state, through different outputs produced by DynaMIT-P.
52
3.5 Representation of ATIS/VMS Scenarios in DynaMIT-P
ATIS strategies (such as diversion strategies) are typically employed due to various stochastic
events (such as incidents) and in order to evaluate them, handling information dissemination at
various levels of sophistication. For example, a Variable Message Sign (VMS) may provide
instantaneous information or it may provide predicted information.
The approach that was
modeled and implemented within DynaMIT to both generates instantaneous/predictive travel
times and to evaluate the network performance is illustrated in the remainder of this section.
DynaMIT was previously using a single network-wide information source to represent in-vehicle
devices providing traffic information and guidance. This framework has been extended to allow
for modeling of VMS devices. Furthermore, these devices have been implemented so that they
may provide link-based or path-based travel time information, as well as either instantaneous, or
predictive information. In this thesis, for development of diversion strategies during incidents,
information to drivers will be disseminated through VMS. A more detailed presentation of this
modeling work can be found in Sundaram (2002).
A brief description on in-vehicle information and VMS representation in DynaMIT-P is
presented next.
In-vehicle Information:
Guided travelers who received in-vehicle information continuously update their paths
using the en-route choice models in DynaMIT-P. The frequency with which drivers update their
paths depends on the nature of information and the frequency with which the information is
updated. Guided travelers use current guidance impedance tables to make en-route decisions.
Currently, in DynaMIT-P in-vehicle drivers receive descriptive information.
VMS Information:
DynaMIT-P models Path-VMS and Link-VMS (see section 3.5.3). Both guided and
unguided drivers may respond to the VMS messages and perform en-route decisions based on
the VMS message. The implemented behavior is shown in Figure 3-4. Guided drivers respond to
53
the VMS signs with probability p1 and ignore with probability 1- p1. Similarly, unguided drivers
respond with probability p2 and ignore with probability 1 - p2. The value of p1 and p2 are inputs
to the model and can be determined through calibration.
Travelers with habitua
U nguidtd
C Guided
p1
S
p2
1 -p1
Follow VMS
Ignore VMS
1 -p 2
I gnore VMS
VM
IFollowO M
Figure 3-4: Travelers Response to VMS
3.5.1 Modeling ATIS/VMS with Instantaneous Information
The time-dependent OD matrix for a time interval is disaggregated to produce a list of travelers.
Based on the information available about the percentage of unguided and guided travelers 2, the
list of travelers is divided into two driver classes:
informed and uninformed drivers.
The
habitual paths of both classes of drivers are obtained by the standard route choice model using
the historical travel times. Uninformed travelers are loaded into the supply with the habitual
2 Guided
drivers are the drivers that have access to in-vehicle information; unguided drivers are the drivers who do
not have access to such information. In the application for the LWC network, all drivers will be assumed to be
unguided.
54
paths and do not make en-route decisions unless they encounter a VMS message.
Informed
vehicles on the other hand may change their routes dynamically based on both in-vehicle and
VMS information.
The instantaneous information is obtained by aggregating travel time information from
the supply, after every information-update period.
For example, if the frequency of the
information updating is 5 minutes, every 5 minutes the travel times obtained from the supply
simulator are broadcast to informed drivers. These informed drivers then make en-route choices
depending on the travel times supplied to them, using appropriate compliance and en-route
choice models. The travel times provided to drivers can be on certain links or on certain paths
depending on the characteristics of the ATIS.
During the supply simulation, whenever any vehicle passes over a link which has a VMS,
a driver may respond to the VMS with certain probabilities.
In this case, traveler behavior
models in the demand component are invoked to simulate en-route choice. The utilities for the
en-route choice models are computed based on the travel times provided by the VMS (for the
specific links or paths) and historical travel times for all the other links. For the instantaneous
information VMS, the frequency of the update of the information that is provided by the VMS is
a design parameter and an input to the system.
3.5.2 Modeling ATIS/VMS with Predictive Information
The important aspect of traffic prediction is the concept of consistency. Based on the guidance
provided, travelers' response to the guidance will influence network conditions and hence the
guidance strategy. Therefore, the problem of obtaining a consistent guidance is a fixed-point
problem. Any ATIS based on predictive guidance has to iterate so that the outcome of the
guidance strategy matches the network conditions after travelers' reactions to the guidance.
An initial set of predicted travel times is assumed (typically this is the historical travel
times) and network conditions are simulated. Based on the simulated travel times, algorithms are
used to obtain a new prediction of travel times, which are used for guidance generation. The
new set of predicted travel times is provided either through in-vehicle or VMS information, and
55
travelers update their paths. If the resulting travel times do not match the predictive travel times
provided to travelers, then consistency has not been achieved and more iteration are necessary.
On convergence, a consistent guidance strategy is obtained.
3.5.3 Dissemination of VMS Information in DynaMIT-P
The information provided through the VMS may be path-based or link-based. Therefore, the
following types of devices can be modeled:
*
Link-VMS: VMS device that displays information on individual links
" Path-VMS: VMS device that displays information on specific paths or sub-paths
Both guided and unguided drivers may respond to the VMS messages and perform enroute decisions based on the provided information. The path travel times for travelers making
en-route decisions in the presence of a VMS are computed based on the following rules.
In the case of a link-VMS, the latest travel times for the links served by the VMS are
substituted for the historical travel times. These latest times are stored in the current guidance
table. Thus, for the example in Figure 3-4, assuming that the VMS is located on link 4 and that it
provides information about link 7, then for the path represented by a sequence of links 6-7-8, the
travel time on link 7 is obtained from the current guidance table and the travel times of all other
links are obtained from the historical travel time table.
F2M
3
VMS
5
9
4
10
-
6
8
7
11
Direction of flow
Figure 3-5: Example to illustrate link-VMS and path-VMS
56
In the case of a path-VMS, while computing path travel times for a particular path, it is
checked whether sub-paths for which the VMS provides guidance is a subset of the current path.
If this is the case, then for all the links in the sub-path the guidance provided by the VMS is
employed to get the link travel times. Otherwise the historical travel times are used for the entire
path. Thus, with reference to Figure 3-4, assuming that the VMS is again located on link 4, and
gives path guidance for path 6-7, then for vehicles following the path 6-7-8 travel times from the
current guidance table are used for links 6 and 7 and the historical travel times are used for all
the other links. However, if for example a vehicle is using the path 4-6-11, then it uses historical
travel times for all the links on this path, since path 4-6-11 does not contain the sub-path 6-7 in
its entirety (even though it contains one of the links that belong in sub-path, i.e. link 6).
The structure of the VMS input file for DynaMIT is explained in Appendix B. The use of
the input file is illustrated through examples.
3.6 Framework of the Planning Tool for the Lower
Westchester County Case Study
The main objective of this thesis is to illustrate the use of a calibrated DynaMIT-P system for
developing/evaluating incident response strategies. The framework described in the following
sections will be used in the chapter 5 for analysis of incident scenarios and the diversion
strategies.
3.6.1 Development of Incident Scenarios
Figure 3-6 describes the step-by-step process for the development of an incident scenario. First
step is to identify and analyze those paths/links that are mostly used by travelers, since the
impact of an incident affecting these paths/links will be more severe. This can be obtained from
the historical data (sensor counts, O-D flows) or from the statistics of segments (densities and
flows) on different paths/links in the Base-Case of DynaMIT-P, which is established during the
calibration process under normal traffic conditions. Next step is to choose a path/link from the
57
set of high occupancy paths/links, for illustrating the impact of an incident. After choosing a high
occupancy path/link, critical location(s) on this path/link will be identified that will have a major
impact on traffic conditions as a result of an incident. These location(s) are found by analyzing
the traffic flows, bottleneck points, number of options for diverting traffic incase of an incident
and historical data about incidents. Incident descriptions will include information on the location,
severity and duration of the incident. Incident details will be fed into the calibrated DynaMIT-P
from the no-incident case, which will estimate the network traffic conditions at different points
of time, assuming travelers follow their habitual routes. Estimated network state will show the
impact of the incident in terms of total number of completed trips, average travel times for
different departure intervals and total vehicle hours.
58
:StartD
Identify and analyze paths and
links with high occupancy
Historical data (flows,
sensor counts, densities)
Choose a path/link
Identify the most critical
location on this path/link
Create the incident at this
critical location
Run calibrated DynaMITP with final planning OD
and equilibrium travel
times
Output (Network
Performace)
-number of complted
trips, travel times etc.
CStop
Figure 3-6: Development of Incident Scenario
59
i t tsoica tion
serity, duration)
severity, duration)
3.6.2 Evaluation of Diversion Strategies
Sundaram [48] has concluded that the Predictive VMS strategy results in the best overall
network performance, both in terms of the number of travelers who complete their trips and the
average travel time. He further mentioned that the Instantaneous VMS strategy might cause
travelers to experience longer travel times, since this strategy does not take into account future
network conditions and may lead to overreaction. This effect may be avoided by using the
predictive VMS scenario with a consistent guidance strategy. Therefore, only predictive VMS
strategy will be applied for evaluating the diversion strategy on the LWC network.
Determining the location of the VMS is an important step in the response strategy. The
VMS strategy in our analysis is constrained by the fact that the location of VMS on the network
is fixed and only these fixed VMS can be used for the diversion of traffic flow. Appropriate
VMS location from these fixed VMS locations will be chosen based on the incident location. In
general, a response strategy designates a set of travel time messages for certain links or paths,
together with the locations of VMS which are fixed over the LWC network. Various measures of
effectiveness obtained during incident scenarios will provide benchmark performance measures
that establish reference points for quantifying the benefits of alternative response strategies.
A framework for off-line generation and evaluation of diversion strategy with
predictive ATIS is shown in Figure 3-7. The first in the process consists of loading the initial
network state and disaggregating the planning OD to produce a list of travelers. An iterative
approach is employed to achieve a consistent guidance. The list of travelers is then loaded into
the supply simulator, which is used for the entire planning horizon to estimate the network
conditions. During the supply simulation, if travelers encounter a VMS, they update their paths
based on en-route choice models. Guided travelers on the other hand continually update their
paths. These travelers perform route choice based on the current guidance table. Network
performance is estimated after all routes are updated and subsequent updates to guidance are
based on aggregated travel times from the supply simulator. A linear combination of the previous
60
current guidance and the latest travel time tables from the supply is used as the guidance
generating algorithm.
Consistency is a key factor in the case of predictive ATIS strategy. Consistency takes into
account the drivers' reaction to the provided information. Therefore, a consistency check is
performed based on the following equation:
L
( Ir t -
It
1=1 t=1
2
<
Ni N,
where TT"' is the link travel time from the guidance table in the current iteration i, on link 1, for
a vehicle with time of entry into the link t minutes after the start of the planning horizon, IT"
is
the corresponding link travel time from the guidance table in the previous iteration i-1, N, is the
number of links in the network, N, is the number of time intervals corresponding to that of the
impedance tables in DynaMIT-P and E is the user defined threshold parameters. If the
convergence criterion is not satisfied, then more iterations are done as indicated in Figure 3-7.
Performance of a response strategy will be evaluated in terms of total savings in travel
times, frequency of travel times, number of completed trips, total vehicle hours etc.
61
Start
Load network at the start
of the planning horizon
Input
Planning OD
Equilibrium Travel Times
-
-
Demand Simulation
Disaggregate Planning OD to creat a list
of travelers
Output
-List of travelers with
habitual behavior
Incident and
VMS
descriptions
I
4i
Supply Simulation
with en-route decisions (In-vehicle /
VMS)
ATIS
Update Guidance
Computed Travel Times
Li
Predictive
Travel Times
Consistency?
No
Yes
Output
-Network Performance
(number of completed
trips, travel times etc.)
Figure 3-7: Evaluation of Diversion Strategy using Predictive ATIS Strategy in DynaMIT-P
62
3.7 Summary
This chapter gave a description of the use of DynaMIT-P for developing incident scenarios and
diversion strategies. First a brief description of DynaMIT was given, followed by the detailed
overview of DynaMIT-P and its main components. Suitability of DynaMIT-P for such planning
purposes was described by its capabilities to do day-to-day and within-day behavior modeling.
Finally, representation of incident scenarios and VMS scenarios in DynaMIT-P were covered
which will be applied during some case studies over the LWC network in the next two chapters.
63
64
Chapter 4
Calibration of DynaMIT-P
Westchester County
for
the
Lower
The objective of this chapter is to calibrate the DynaMIT-P for the Lower Westchester County
(LWC), using the methodology described in the chapter 2. Calibration of the DynaMIT-P will
establish the no-incident case, which will be used for comparing incident scenarios and diversion
strategies in the next chapter.
A brief description of the study network is given first. The data available from various
sources and issues associated with data are discussed. Critical calibration results, which are to be
obtained before the use of DynaMIT-P for incident scenario and diversion strategy, are presented
next.
4.1 The Lower Westchester County (LWC) Dataset
The data used in this research was collected from LWC, New York, USA. Descriptions of the
main features of the network, available data, issues and assumptions associated with the data are
given in the following sections.
4.1.1 Network Description
The primary network of the project is the Lower Westchester County (LWC) freeways and
parkways network. This network is just north of Manhattan, New York. It thus has heavy traffic
on some of the freeways and parkways, resulting from a varied mix of commuters and travelers.
Freeways and Parkways in the focus network include 1-87 (the New York State Thruway), 1-95
65
(the New England Thruway), 1-287 (the Cross Westchester Expressway), 1-684, the Cross
County Parkway, the Hutchinson River Parkway, the Sprain Brook Parkway, the Saw Mill River
Parkway, the Bronx River Parkway and the Taconic State Parkway. Four adjoining arterials
(Tuchahoe Road, Ardsley Road, Hartsdale Road and Weaver Street) and Routes 9, 100 22, and
119 were also added to capture the effect of any diversion strategies in case of incident. Most of
the freeways and parkways, except Cross County Parkway and 1-287, run south- north, and vice
versa, direction. The general extent of the project network is shown in Figure 4-1.
Figure 4-1: The Lower Westchester County Network
The network is represented as a set of 579 O-D pairs connected by 1659 directed links.
These links represent the physical links on the network, and are further subdivided into 2421
segments to model changing link characteristics.
66
4.1.2 Data Description and Analysis
This section describes the available Origin-Destination data and traffic sensor data that are
required for the calibration.
4.1.2.1 Origin - Destination Data
The New York Metropolitan Transportation Council (NYMTC) data included demand
information in the form of very detailed origin-destination (OD) flows. Morning peak (6am10am) data was used for the analysis.
The data was not suitable for the purposes of this application. In particular, many of the
OD flows were not contributing to traffic in the study area. Furthermore, many origins and
destinations corresponded to very small areas, thus leading into very small demand flows.
Dealing with very small demands in demand estimation and prediction can be problematic.
NYMTC demand information was analyzed and the origins and destinations were
aggregated to a manageable set of zones. The total number of OD pairs was thus reduced to
3344. Further analysis of this demand table indicated that most of the OD pairs had demand
lower than 100 vehicles (for the entire peak period, i.e. less than 6 vehicles per 15 minute
interval). In order to avoid numerical stability issues during the OD estimation and prediction,
we decided to further aggregate OD flows. The resulting demand table consists of 579 OD pairs.
Figure 4-2 indicates the O-D pairs for LWC network. The thickness of the lines
connecting the origin and destination is a measure of the magnitude of the O-D flow.
67
Figure 4-2: O-D Pairs for LWC Network
4.1.2.2 Surveillance Data
This section describes the available surveillance data and how it was processed for the purposes
of this calibration effort.
Aggregate Archived Data
Aggregate data from 120 sensors were originally provided by NYSDOT. This was later
supplemented by data from 12 more sensors, leading to a total of 132 sensors' data. The sensors
were located on 4 freeways and 6 parkways in the LWC network. These freeways and parkways
include 1-87 (the New York State Thruway), 1-95 (the New England Thruway), 1-287 (the Cross
Westchester Expressway), 1-684, the Cross County Parkway, the Hutchinson River Parkway, the
Bear Mountain Parkway, the Sprain Brook Parkway, the Saw Mill River Parkway and the
Taconic State Parkway. No surveillance information was available for several road arteries in the
68
study network, namely the Bronx River Parkway, 4 adjoining arterials (Tuchahoe Road, Ardsley
Road, Hartsdale Road and Weaver Street), Routes 9, 100 and 119. The original data was already
aggregated in hourly intervals, and up to three (different) days of data were available for each
sensor location.
Table 4-1 presents in detail the spatial distribution of the sensors in the network, as well
as the period for which data is available from each sensor. 60 sensors are located on Parkways
and the rest of the 72 sensors are located on the parkways. Out of a total of 132 sensors, data
from 94 sensors (71%) were collected during 2000, data from 12 sensors (9%) were collected
during 2001 and data from the remaining 26 sensors (20%) are from 2002.
Analysis of the available archived sensor data indicated that data from only 58 out of the
132 sensors could be used for the purposes of this study. The remaining detectors were either
situated outside the study area, or provided inconsistent data. Table 4-1 shows the distribution of
these finally selected sensors on freeways and parkways. 32 sensors are located on freeways
while the remaining 26 sensors are located on parkways. Out of 58 usable sensors, 42 sensors'
data (73%) were collected during 2000, 2 sensors' data (3%) were collected during 2001 and the
remaining 14 sensors (24%) were collected during 2002. The location of sensors in the final
network is shown in Figure 4-3. Furthermore, a detailed analysis of the traffic sensors' data is
provided in Appendix C.
69
Time Horizon for All Sensors Data
Total
Freeway Sensors
099 9
0 0
L <
ss
O
9
1-87/NYS Thruway
14
4
1-95
18
12
1-287/Cross Westchester Express
14
8
1-684
14
Total Freeway sensors
60
0
u''
z0
9
< 2
Collection
JC
<
.0 Z
a n
LL
<
2
n'5
0
<
2
8
6
4
12
2
2
Parkway Sensors
Saw Mill River Parkway
14
Bear Mountain Parkway
4
Cross Country Parkway
2
Taconic State Parkway
18
Sprain Brook Parkway
12
Hutchinson River Parkway
22
Total Parkway Sensors
72
Total Sensors
6
6
2
2
2
2
18
6
12
1
2
2 #
_
_
4
8
132
Time Horizon for All Sensors Data Collection
Freeway Sensors
1-87/NYS Thruway
1-95
1-287/Cross Westchester Express
1-684
11
16
Total Freeway sensors
32
Parkway Sensors
Saw Mill River Parkway
Sprain Brook Parkway
2
Hutchinson River Parkway
16
Total Parkway Sensors
26
Total Sensors
000
CO4)C
.
0.
<
(
0
09
0009990
3Q)
OO
z0
0c
u1
0
:5
<
1
3
21
1
1
2
1
11
Q
a
7
4
1
4
0l
~' ~'
4
12
1
Taconic State Parkway
Cross Country Parkway
0
4
7
0
0
1
Bear Mountain Parkway
99
Total 90
Sensors -
3
58
Table 4-1: Total (top) and Usable (bottom) Sensors Received from NYSDOT
70
7
16
?~S'~
/
a
V
870
7
87021
Figure 4-3: Location of Sensors on the LWC Network
71
72
Figures 4-4 and 4-5 show examples of freeway sensors count variability and
Figure 4-6 shows parkway sensor count variability. Some of the conclusions about the
traffic pattern can easily be drawn from these graphs. The maximum count reported by
sensors, during peak hours, on most of the freeways is between 5000 to 6000 vehicles per
hour. Maximum count reported by most of the parkways is between 3000 to 6000
vehicles per hour, depending on their location on the network. Most of the sensors report
morning peak between 7 AM to 9 AM and evening peak between 4 PM to 6 PM. Figure
4-4 clearly indicates that morning peak sensor count on 1-95 in North-bound direction is
much higher than the South-bound direction count, and vice versa for evening peak hour.
It is also observed that sensor counts never go below 1000 vehicles per hour at this
section of 1-95, making 1-95 as a high traffic volume part of the network. Figure 4-5
shows the distribution of sensor counts in East-bound and West-bound direction, for a
sensor on 1-287. Unlike the 1-95 sensor discussed above, there is not much of a difference
in the morning and evening traffic volume at that section, in both the direction. Figure 46 shows a sensor count distribution on Saw Mill River Parkway, which shows that
evening peak has more traffic in North-bound direction compared to morning and vice
versa for South-bound direction traffic.
73
ID: 870213 - Secton From Inter 15 Rt I To: Inter16 North Ave Cedar St -(NB)
Sensor Count Variation
moo.
4000
--
-e-
Monday
11il0o
Tuesdy
1117100
We
\h
dayI 1/80
day 11i/90
2000
1000
0
4
a
20
12
24
Timne of Day
ID: 870213 - Section: F mon Inter 15 Rt I To: Inter 16 North Ave Cedar St- (SB)
Sensor Count Vafation
4000
-+MOndy 11i00
S- Tues day 1100
00D
:3000
3
-
Thurs day I1119W
1000
0
0
4
8
11
12
20
24
Time of Day
Figure 4-4: Sensor Count Variation for a Freeway (1-95, North and South Bound) Sensor
74
FrmAcc Hukt~son River PkvW To: Ac Rt 120- JEB)
U: 870739 -Seclorc:
SensorCount Vadation
8--$-Moniday
b
.--
>
-
I ffl
Tmesday 111ThOO
e0e~
r
24
16
812
04
Time of Day
ID' 870T39-Section Frorm Acc Hutchison River Pkwy To: Acc Rt 120 - (WB)
Sensor Count Vanation
-dyIO
2'-~
A-
4W -
~--Tmsday 1117,00
Weurmlay
1/00
J
0-
4
16
812
24
Time of 03y
Figure 4-5: Sensor Count Variation for a Freeway (1-287, East and West Bound) Sensor
75
...........
ID: 870914 - Seclon: From: Int Old Saw Ml Rs To: Acc Rt9A -(NB)
Sensor Count Variation
3m
3000
Moo
--
Manday 7/31 /00
Tuesday 10/2400
--
Wednesday 10I200
ThLrsday 10/23'100
S1500
1000
500
0
4
0
8
18
12
2D
24
Time of Day
ID: 870914 - Section: From: Int Old Saw 111 Rs To: Acc Rt 9A -(SB)
Sensor Count Vanation
-
Monday 7/31/00
-- Tuesday 10/24M0
Wednesday 10/M
Thursday 10/2/00
1500
100--
0
04
1e
812
20
24
Time of Day
Figure 4-6: Sensor Count Variation for a Parkway (Saw Mill River Parkway, North and
South Bound) Sensor
76
Issues and Assumptions Regarding Traffic Sensor Data
There were some important issues associated with traffic sensors data received from
NYSDOT. Due to unavailability of TRANSMIT data from NYSDOT, loop detectors
(sensor counts) data collected at different points over LWC network was provided by the
NYSDOT, for the purpose of calibration and response strategies. All the sensor counts
are from weekdays aggregated for a period of 1 hour. Most of the sensors have at most 3
days of data but sensor counts for different sensors have been recorded on different dates.
Therefore, there were two major challenges for incorporating these sensors counts
into DynaMIT. First, DynaMIT needs sensor counts for finer interval for its estimation
and prediction models. Unavailability of finer counts lead us to make required changes in
DynaMIT so that it can take into account hourly sensor counts for its estimation and
prediction.
Second, since sensor counts are from different dates and years (Table 4-1), this
will again require a methodology for pooling the data. This was the bigger task, as for
good calibration results we require traffic sensor data on similar time periods. As a result
of this, a traffic sensor count variation was observed and studied at different locations
over the network. It was important to see if there are traffic demand pattern changes over
the three years (2000, 2001 and 2002), and how to account for these changes if they are
present. As a result of this analysis it was found that Wednesday is the only common
weekday where all the sensors have data, although for different years and dates as said
above. We decided to look at sensors data on this day only as it will provide better
comparisons for traffic demand pattern. Some of the sensors that were showing major
differences in the sensor count value as well as variation were removed from the analysis,
as it will be difficult to capture these differences over the years when we have only one
weekday of data for calibration purposes. As said above, after these analysis only 58
sensors out of 132 sensors reported usable data. Although these 58 sensors were from
different years, it was observed that traffic demand pattern is comparable for calibration
purposes. As a result, it was decided to pool all the different Wednesday sensors data
together and to consider it as a single day of sensor data. This is a very crude assumption
but nothing else could be done, given the quality and availability of data. Moreover,
77
changes in the traffic demand pattern were not very significant for 58 usable sensors. It
reinforces our assumption of pooling different Wednesday data into a single day.
Acoustic Sensor Data
In order to estimate various supply simulator parameters (various speed-density
relationship parameters like, jam density, parameter cc and P, minimum speed, free flow
speed), we need to have a detailed traffic sensor data that also have occupancy and speed
measurements, other than just counts. Estimation of these supply parameters is very
critical in order to replicate the field traffic conditions. Although we need detailed traffic
sensors data at different locations over the network for estimating supply parameters for
different segments, only one detailed traffic sensor data was provided by NYSDOT. This
sensor is located near Exit 5 on 1-287 Eastbound freeway and it uses acoustic
relationships to calculate counts, occupancy and speed of the vehicle. Many days of data
(including weekdays and weekends) from the month of August and November in year
2002 were provided by the NYSDOT. This acoustic sensor reports enormous amount of
data at random intervals like 7, 11, 23 seconds etc.
Sensor data was processed and the graph for the day of November 8 (Friday),
2002 is shown in the Figure 4-7. There are 3 lanes at that section of the freeway. Flow is
calculated by aggregating one-minute counts and scaling it to hourly value. Average
speed for that minute is calculated by dividing the product of counts and speed by total
number of vehicles reported during that minute. Classical flow-speed-density relationship
(flow = density x speed) has been used to get an idea about the density variation. Figure
4-8 shows the aggregated hourly sensor counts variation with the time of the day. It is
clear from the Figure 4-7 and 4-8 that values reported by the sensor are erroneous. As
you can see from the Figure 4-8, hourly counts (10,000 - 12,000 per hour) reported at
that location is much higher than the expected value for 3 lanes. Figure 4-7 shows the
counterintuitive speed-density and speed-flow diagram. Problems with data quality were
discussed with NYSDOT, who acknowledge that the data reported by the acoustic sensor
is not correct.
Due to this problem the acoustic sensor data is not been used for estimating
supply simulator parameters. Instead, some default widely accepted values of these
78
parameters were tried, and finally supply parameters value from Irvine, CA network were
used. Irvine network is an urban network consisting of freeways and arterials, which is
similar to LWC network.
79
November 8, 2002, Friday
Flow Vs Speed Variaton
8070 -
60
-
JL40
.20
10
0
0
1000
2000
3000
4000
!000
000
Q. Flow (vehicles per hour per lane)
7000
8000
K ovember 8, 2002, F niday
Speed Vs Density Vaniation
70
m.50
.
10
0
20
40
K
80
80
100
120
140
120
140
Density. (vehicles per mile per lane)
November 8,2002, F niday
Flow Vs Density Variation
8000 70008000 U-
5000 -
EL
4000
-
2000
-
1000
-
7Z
EL
0
4
0
20
80
100
40
50
K. Density. (vehic lesper mile per lane)
Figure 4-7: Speed-Density-Flow Variation from Acoustic Sensor Data, located near Exit5 on 1-287 Eastbound Freeway
80
Hourly Sensor Counts Variation
(Nov 8, 2002, Friday)
14000
12000 10000
8000
6000
5
.
4000
2000.
0
0:00:00
200:00
4:00:00
6f00:00
8:00:00 10:00:00 1200:00 14:00:00 16:00:00
Time of Day
Figure 4-8: Hourly Sensor Count Variation
4.1.3 Network Data and Network Preparation
This section focuses on the available data for network generation. The network
representation of the study area was based on geographic information from New York
Metropolitan Transportation Council (NYMTC), web-based resources, and additional
information and feedback provided by NYSDOT. These data sources are described in the
following subsections.
4.1.3.1 NYMTC Data
A large amount of data was provided to MIT in 2000. This data was not suitable for the
purposes of ITS applications and substantial work was performed to process this data.
The main issues were that the data covered many areas outside the scope of the current
81
project and also included irrelevant information. The data was processed and reduced to
the appropriate scope and level of detail.
While the NYMTC data was complete, they were combined with U.S. Census
TIGER files and geographic files provided with the TransCAD software. The additional
data sources were helpful as references and mapping tools.
Figure 4-9: NYMTC Representation of Interchange of 1-87 with Cross County
Parkway
One major shortcoming of these data pertains to the detail with which the network
is represented. In particular, since this data is usually used for planning applications, the
network is represented as a collection of links and nodes. As shown above (Figure 4-10),
however, the mesoscopic models within DynaMIT's supply simulator require a more
detailed representation of the network topology. The next subsection describes additional
sources of geographical data.
4.1.3.2 Additional Geographical Data
As mentioned above, the NYMTC data does not provide the necessary detail for the
network representation in DynaMIT. Therefore, additional information sources were
sought. From the publicly available sources, we selected the following two:
82
" New York State GIS Clearinghouse (http://www.nysgis.state.ny.us/)
" GlobeXplorer (http://www.globexplorer.com)
These two resources provide detailed ortho-photos that include the study area. This
information was used to verify the topology of some complicated interchanges, as well as
details not available in the NYMTC data (including lane connectivity, ramp topology,
and curvature; see Figure 4-11).
Figure 4-10: Ortho-picture of Interchange of 1-87 with Cross County Parkway
from GlobeXplorer
4.1.3.3 Network Generation
One of the most important inputs to DynaMIT is the network representation of the study
area. The available data sources have already been presented in Section 4.1.3.1 and
83
Section 4.1.3.2. Furthermore, a number of software tools have been used in the
generation of the network representation:
* Caliper's TransCAD was used for viewing and manipulation of the original
NYMTC GIS files. Information was exported from TransCAD in various formats
for further processing in other tools.
* Microsoft Excel was used for the analysis of geographic data exported from
TransCAD.
* RNE, the graphical network editor developed at MIT ITS Program, was used for
the generation of the final network file in DynaMIT format.
The first step was the generation of a skeleton network file in TransCAD. Starting
from the raw NYMTC data set, all links that were outside of the study area were
removed. Furthermore, local streets that are not being modeled were also removed. The
resulting network captured most of the traffic network of the study area but was still not
complete and lacked the level of detail required by DynaMIT.
The data was exported in a suitable format and then imported to Excel for further
processing. Links in the original NYMTC data were coded as bi-directional links.
DynaMIT, however, requires that each bi-directional link is substituted by two singledirectional links. Furthermore, some information on some of the links was missing.
Using detailed information from the GIS resources available from the web (see
Section 4.3.1.1), the missing information was completed. The skeleton network
representation was then converted to the DynaMIT network file format using a software
utility that was developed for this task.
The final stage of the network coding was performed using the graphical road
network editor. Additional details of the network representation were added to the
network. In particular, interchange representation was considerably improved, lane
connections were generated, and segment curvatures were introduced (Figure 4-12). (All
this information was not present in the NYMTC data). The completed network was
presented to NYSDOT and feedback was obtained. As a result, an additional number of
arteries were added. Furthermore, the first simulation using this network indicated that
there was a lack of horizontal arteries. In cooperation with NYSDOT a number of such
arteries were identified and added.
84
Final network details used for calibration of DynaMIT are:
" Number of Links = 1659
" Number of Segments = 2421
*
Number of lanes = 5242
" Number of OD Pairs = 579
*
Number of Sensors = 58
85
TransCAD Image of Interchange of I-
GlobeXplorer Image of Interchange of
87 with Cross County Parkwa
1-87 with Cross County Parkway
Screen-shot of the same Interchange in Graphical Road Network Editor
(RNE)
Figure 4-11: Example of using TransCAD and GlobeXplorer Image to Construct Detailed
Interchange required for the purpose of Traffic Modeling in DynaMIT.
86
4.2 Calibration Results
Before DynaMIT-P is employed, a critical task is the calibration of the system. The
Following section covers the calibration results. The input files required for DynaMIT-P
run and calibration are described in Appendix A.
4.2.1 Supply Simulator Parameters
As mentioned above, the supply parameters that need to be calibrated are the:
*
capacities of the network segments, and
" parameters of the speed-density relationships.
The values for the capacities of the network segments have been computed based on
the Highway Capacity Manual (Reference [11]). In particular, the resulting values for the
freeways are 2200 vphpl (vehicles per hour per lane), while for the arterials (Tuchahoe
Road, Ardsley Road, Hartsdale Road and Weaver Street) and routes 9, 22, and 100, was
set to 1900 vphpl. Ramp capacities were set to 1600 vphpl or 1800 vphpl depending on
the location and geometry of the ramp.
Detailed speed and occupancy data are required for the calibration of speed-density
relationship parameters. From the available data, only the acoustic sensor data was
conceptually appropriate for this purpose. However, analysis of this data provided
counterintuitive results, thus indicating that the data is not very reliable. In particular, the
measurements obtained from these sensors violated the fundamental traffic principles.
These issues have been demonstrated in Section 4.1.2.2.
As a result, it was necessary to transfer values from other applications. This is
admittedly a difficult exercise and should always be performed with extreme caution.
Based on previous calibration experience from similar networks (e.g. the Irvine, CA,
network, which is also a very congested, suburban network composed of freeway and
arterial routes) appropriate values were selected for the model parameters. Extensive tests
were performed to ensure that the parameters adequately capture prevailing traffic
conditions.
87
The speed-density relationship parameters are shown in the following table. Different
sets of parameters were selected for freeways, arterials and ramps.
Minimum
Density
Free
Flow
Minimum
Speed
Jam
Density
Speed
Umin
kjam
Uf
(mph)
(vpmpl)
(vpmpl)
(mph)
Parameter
Parameter
a
Freeways
65
20
0.125
0.015625
2.0
0.5
Arterials
50
20
0.0625
0.00625
1.5
0.32
Ramps
45
20
0.125
0.0125
1.5
0.3
Table 4-2: Speed Density Relationship Parameters per Road Type.
where,
mph = miles per hour,
vpmpl = vehicles per meter per lane
88
4.2.2 Demand Simulator Parameters
4.2.2.1 Defining the Period of Study and Generating Seed O-D Flows
The period of study was defined as the AM peak, from 6:00 AM to 10:00 AM3 . However,
it had to be ensured that the first interval we estimated should not start with empty
network. As a result few runs of DynaMIT were made for 4 AM 4 to 6 AM interval in
order to have a starting network state, loaded with vehicles at 6 AM when we start doing
our calibration. The study interval was divided into equal subintervals of duration of 60
minutes. This discretization was based on the fact that we have only 60 minutes sensor
counts data, and finer sensor data was unavailable. A reasonable starting estimate of the
O-D flows was constructed manually to start the O-D estimation process. The planning
matrix of static O-D flows of 4 hours duration (this static O-D flows from 6 AM to 10
AM was based on the NYMTC data) was distributed across the time intervals within the
study period. The fraction of flow assigned to a particular interval was proportional to a
factor that was computed for each sensor as the ratio of the counts measured in the
current interval to that measured across the peak period (6 AM - 10 AM). This factor for
each O-D flow was approximated by considering an average of the primary sensors that
measured the O-D flow under study.
4.2.2.2 Parameters for Path Choice Set Generation
An important step in the calibration process was the generation of a good set of paths for
each O-D pair of interest. Optimum parameters for the path generation algorithm were
identified so as to capture most of the feasible paths for every O-D pair. A suitable path
set was obtained using 40 random draws to complement the set of link-elimination-based
shortest paths from every link to a destination node. Recognizing the need to replicate
3Flows were found to drop beyond 10:00 AM.
4 Sensors indicated that network was predominantly empty at 4:00 AM.
89
most of the freeway-based paths, an internal freeway "bias" of 0.6 was used to force the
path generation algorithm into preferring paths with longer freeway sections. The random
draws helped augmenting this set with arterial paths. Manual inspection confirmed that
most of the practical alternatives had been selected in the path generation stage. The final
set contained a total of 7541 paths between 579 origin-destination paths.
4.2.2.3 Simplifying Assumptions
Several assumptions were made in order to accommodate practical considerations while
estimating the model parameters with the limited data available. The error covariance
matrices Vh and Wh (section 3.2.1.1) were assumed to possess a diagonal structure, in
order to ensure that we had enough observations to estimate the elements of these
matrices from just one day of data. The structure of the error covariances was also
assumed to remain constant across the peak period. The autoregressive factor was
assumed to be of degree one, meaning that the deviation of flow between O-D pair r from
their historical values does not depend on the prior interval flow deviations between the
O-D pair r. The O-D flows themselves were further grouped into high and low categories,
based on the magnitude of the OD flows that they represented.
4.2.2.4 Error Statistics
The following error statistics were used in analyzing the results:
1. Root Mean Square (RMS) Error =
Fi
N
-
2. Root Mean Square Normalized (RMSN) Error =
where,
90
N = number of observations,
y = observed value,
9,= estimated value,
4.2.2.5 O-D Estimation and Route Choice Parameters Values
As mentioned during the surveillance data analysis (section 4.1.2.2), data collected
during Wednesdays was used for the calibration. The seed O-D flows computed from the
static planning matrix were used as the input for the O-D estimation module. In the
absence of any initial error covariance estimates, a weighted least squares approach was
adopted, with the sensor counts being assigned higher weights than the target O-D flows.
This step was employed in order to extract all the O-D flows information from the sensor
counts, while obtaining structural information about the overall demand distribution from
the seed O-D flows. The weights used for the O-D flows were not kept constant across
the O-D pairs. Low flows (as indicated by the planning matrix) were assigned higher
weights in order to maintain them in an acceptable range, and capture the fact that such
flows normally exhibit low variability. As pointed in section 3.2.1.1, flows for each time
interval were estimated sequentially using the GLS formulation.
As discussed in the chapters 2 and 3 (Figure 2-3 and Figure 3-3), problem of OD
estimation is a fixed-point problem. The solution of OD estimation is obtained by an
iterative process between the OD estimation algorithm and the supply simulator, linked
by an assignment matrix. The planning OD was updated based on the framework shown
in Figure 3-3. Several iterations of OD estimation and equilibrium travel times
computation were performed until the overall convergence criterion was satisfied. At the
end of this procedure, we had a set of equilibrium travel times and the final estimated
OD.
The computation of the equilibrium travel times and the results obtained are
described with respect to some important OD pairs for the LWC network. Figure 4-13
illustrates these origin and destination nodes along with a sample of paths in the path
choice set of the OD pair 4-6. Origin nodes are numbered as 1, 2, 3 and 4, and destination
91
nodes are numbered as 5, 6, and 7. Origin nodes 1, 2, 3 and 4 are grouped together for the
analysis; since these nodes have similar characteristics in terms of demand, travel times,
number of paths, etc. Figure 4-14 shows the equilibrium travel times (after convergence)
for these OD pairs, during the peak demand period from 7:00 AM to 10:00 AM.
4
&Qfl
Rine
S
H
Cros CouatyPkwy
Rckt
o
k
~ 4e
Figure 4-12: Network showing Origin Nodes (1,2,3,4) and Destination Nodes (5,6,7);
Examples of Paths for the OD pair 4-6.
92
w
4500
4000
3500
20
1000
6:00 6:15 6:30 6:45 7:00
7:15
7:30 7:45
8:00
8:15 8:30
8:45 990
9-15
930 9:45 10:00
Travel Time (seconds)
Figure 4-13: Travelers' Equilibrium Travel Times (after convergence) between all OD
Pairs "1, 2, 3, 4 - 5, 6, 7".
Figure 4-15 shows another set of OD pairs ("1, 2, 3, 4, 8 - 9, 10, 11") that were
considered for the equilibrium travel times computation analysis, along with the
examples of paths between these OD pairs. Origin nodes are numbered as 1, 2, 3, 4 and 8,
and destination nodes are numbered as 9, 10, and 11. Origin nodes 1, 2, 3, 4 and 8 are
grouped together for the analysis purpose; since these nodes have similar characteristics
in terms of demand, travel times, number of paths, etc. Figure 4-16 shows equilibrium
travel times for east-bound travelers between these OD pairs, during the period of 7:00
AM to 9:00 AM.
93
04
Rt 9
Taconi 'tate Pkwy
8
10
SprainyBrook Pkwy
Saw Ma River Pkw
14~~
Bronx River Pkwyl
Figure 4-14: Network showing Origin Nodes (1,2,3,4,8) and Destination Nodes (9,10,11);
Examples of Paths for the OD pairs is marked with dark color.
4000
3500
3000
2500
1
S20DO
15DO
1000
So-
0
6:00
6:15
6:30
6:45
7:00
7:15
7:30
7:45
880
8:15
830
8:45
Departure Time
Figure 4-15: Travelers' Equilibrium Travel Times (after convergence) between all OD
Pairs "1, 2, 3, 4, 8 - 9, 10, 11".
94
9:00
Several iterations were performed in conjunction with the O-D estimation step to
determine an optimal set of parameters for the route choice model. Different combination
of parameters 1,and $2 (see section 3.2.1.1 and equation (3.3) for details; $1= coefficient
of arterial travel time, $2= freeway bias 5) were evaluated, covering a wide range of route
choice situations. One would expect
1 to have a negative sign and
2 to
be a positive
number between 0 and 1. The optimal values for ($1, $2) were found to be (-0.025, 0.80).
4.2.2.6 Sensor Counts Comparisons
Figures 4-17 to 4-20 graphically depict the comparison of estimated (simulated) and
observed (field) sensor counts during the different intervals in the peak period (6 AM 10 AM). These graphs represent the final estimated counts when convergence has been
achieved (after the final iteration of the calibration phase for each time interval). Counts
in the graph represent the hourly counts per lane. In an ideal situation all points in these
figures (representing individual sensor measurements) would fall on the "45 degree line",
indicated by a thin line in the figures. This line would indicate perfect match between the
observed and estimated counts. Furthermore, Table 4-3 shows the final Root Mean
Square (RMS) and Root Mean Square Normalized (RMSN) error values after
convergence, for each time interval.
5 Freeway bias attempts to capture the driver's preference for a freeway section over an arterial section.
95
-
~LE~UL
-
ar-
- --
-
-
-G-
2 0 00
S
1500
EM100
w
500
0
500
1000
1500
2000
2500
3000
Observed Counts at 7 AM
Figure 4-16: Estimated vs. Observed Sensors Count from 6:00 AM to 7:00 AM
3000 --
----
2500
2000
S1500
1000
a,
:500
0
0
500
1000
1500
2000
2500
3000
Observed Counts at 8 AM
Figure 4-17: Estimated vs. Observed Sensors Count from 7:00 AM to 8:00 AM
96
3000
i ka
WA
2500
S2000
1500
1000
500
0
500
1000
1500
2000
2500
3000
Observed Counts at 9 AM
Figure 4-18: Estimated vs. Observed Sensors Count from 8:00 AM to 9:00 AM
3000
25DO
S 2000
1000
500
0
0
500
1000
1500
2000
2500
3000
Obseived Counts at 10 AM
Figure 4-19: Estimated vs. Observed Sensors Count from 9:00 AM to 10:00 AM
97
Time Interval
RMS
RMSN
6:00 AM -7:00 AM
73.2645
19.1538
7:00 AM -8:00 AM
105.3485
22.7966
8:00 AM - 9:00 AM
145.2776
31.5403
9:00 AM - 10:00 AM
103.9146
25.1304
Table 4-3: Final RMS and RMSN Error for Each Time Interval
4.4 Summary
A brief description of the study network was given first. The data available from various
sources and issues associated with data were discussed. Calibration results of the
DynaMIT-P for the Lower Westchester County (LWC) were presented, using the
methodology described in Chapter 2. Calibration of the DynaMIT-P established two noincident cases, which will be used for comparing incident scenarios and diversion
strategies in the next chapter.
98
Chapter 5
Case Study
The objective of this chapter is to implement the framework described in chapter 3 for
incident scenarios and diversion strategies, using traffic surveillance data obtained from
NYSDOT for the Lower Westchester County (LWC) network. The Incident and VMS
scenarios are evaluated based on a hypothetical incident on the LWC network. Scenario 1
focuses on the south-bound direction traffic while Scenario 2 focuses on the east-bound
direction traffic between major OD pairs. DynaMIT-P has been used to evaluate the
predictive information through the VMS.
5.1 Scenario 1
A peak demand period between 7:00 AM and 10:00 AM was chosen for the scenario 1.
5.1.1 No Incident
The experienced travel times and the results obtained are described with respect to some
important OD pairs for the LWC network. Figure 5-1 illustrates these origin and
destination nodes along with a sample of paths in the path choice set between OD pair 46. Origin nodes are numbered as 1, 2, 3 and 4, and destination nodes are numbered as 5,
6, and 7. Origin nodes 1, 2, 3 and 4 are grouped together for the analysis, since these
nodes have similar characteristics in terms of demand, travel times, number of paths, etc.,
and also because demand between a single OD pair is not significant enough to observe
the impact of an incident and VMS messages. Figure 5-2 shows the experienced travel
times for these OD pairs ("1, 2, 3, 4 - 5, 6, 7).
99
-----
4
2
Tacon Smekwy
Sprain $ok Pkwy
T-297
R 00
Bro
Rier PkMc
0A ?VoPkWy
22
Cross County
5
7
Figure 5-1: Network showing Origin Nodes (1,2,3,4) and Destination Nodes (5,6,7);
Examples of Paths for the OD pair 4-6.
5000
4500
3500
25D
00
250
1000
SOD
0
6:00 6:15 6:30
6:45 7:00
7:15
7:30
7:45
8:00
8:15
8:30
8:45 90D
9:15 9:30 9:45 10:00
Tmvel Time (seconds)
Figure 5-2: Travelers' Experienced Travel Times (after convergence) between all OD
Pairs "1, 2, 3, 4 - 5, 6, 7".
100
Figure 5-3 shows that most of the travelers experienced travel times between 1000
and 2000 seconds. Only few travelers experienced low travel times (less than 1000
seconds) and high travel times (more than 2500 seconds). The average travel time
experienced at different time intervals is shown in Figure 5-4. As can be seen from the
figure, average travel time between OD pairs "1, 2, 3, 4 - 5, 6, 7" increases from 1000 1500 seconds (non-congestion time) to 2000 - 2500 seconds (congestion time). For these
OD pairs average travel time starts decreasing after 8:45 AM.
3500
3000
2500
2000
~1500
500
0
<= 1000
10001500
1500 2000
2000
2500
25003000
30003500
35004000
40004500
45005000
Travel Tine (seconds)
Figure 5-3: Frequency of Experienced Travel Times for all OD Pairs "1, 2, 3, 4 - 5, 6, 7".
101
2500 &
C
i
2000
<=7:15 - 7:30- 7:45 - 8:00 - 8:15- 8:30 - 8:45- 9:00 - 9:15 - 9:30 - 9:45 7:15 7:20 7:45 8:00 8:15 8 :30 B:45 9:00 9:15 9:20 9:45 10:00
Departure Tin* Intderval
Figure 5-4: Average Travel Times between OD Pairs "1, 2, 3, 4 - 5, 6, 7".
5.1.2 Incident without VMS
This section presents the impact of an incident on the network. The planning OD and the
equilibrium travel times obtained from the no-incident-case during the time period 7:00
AM to 10:00 AM were used as inputs for the scenario evaluation. The scenario that is
discussed in this section is based on the impact of a hypothetical incident. An incident is
introduced in the system at the location shown in Figure 5-5. This location is on Saw Mill
River Parkway in south-bound direction. This Parkway along with Sprain Brook Parkway
and Taconic State Parkway is highly used by the vehicles. The time of the incident is
from 7:45 AM to 8:50 AM and twenty percent of the segment capacity is available during
the incident. This kind of severity of incident is common in this network where a minor
incident lasts for about half an hour to one hour. Major incidents may last for 3-4 hours
where all the lanes are blocked for the traffic. The analysis focuses on OD pairs "1, 2, 3,
4 - 5, 6, 7", which are the primary OD pairs affected by the incident. Figure 5-5 also
shows some examples of paths associated with OD pair 4-6 in dark color.
102
-
-
--
-
4
2
7,45 AM to 9:50 AM
9Tinte:
Rt,
L4cation Of Incident
Severity: 80 % capacity reductian
Tacomc $tate Pkwy,
1-97
sprasti Brook Pkwy
$SawMIII
Rvor
Rt 1,00
fNWf
Brort Rivmr Tkwy
Hchink
trsky
Cross County
56
7
Figure 5-5: Location of the Incident 1 and Examples of Paths for the OD pair 4-6.
5.1.2.1 Impact of the Incident
To illustrate the impact of the incident alone, the incident was introduced in the system
without VMS and the network performance was analyzed assuming that travelers follow
their habitual routes (established by the equilibrium process). The resulting plot of the
travel times experienced by the travelers based on their departure time is shown in Figure
5-6. Impact of incident is quite obvious as travel times during the peak period increase
from 3000 seconds (no-incident case) to 4500 seconds, an increase of 150 percent.
103
As pointed in the last chapter, origin and destination nodes are grouped together
for the analysis, since these nodes have similar characteristics in terms of demand, travel
times, number of paths, etc., and also because demand between a single OD pair is not
significant enough to observe the impact of an incident and VMS messages. In order to
illustrate the impact of the incident and as a reference for future comparisons, the
frequency of experienced travel times is further analyzed based on the aggregated OD
pairs, as done in the last chapter. Figures 5-7 shows the frequency of experienced travel
times for aggregated OD pairs. Figure 5-7 clearly depicts that as a result of the incident,
some of the travelers experienced high travel times of 3000 - 4500 seconds, which was
not observed in the no-incident case. Number of travelers having travel times in the range
1000 - 2500 seconds has reduced significantly and those with travel times in the range
2500 - 3000 seconds have increased significantly.
5000
4500
4000
3500
3000
2500
20DO
1500
1000
50
6-00
6:15
6:30 6:45 7:00
7:15 730 7:45
8:00
8:15
8:30 8:45
900
9:15 9:30 9:45 1000
Figure 5-6: Impact of Incident 1. Travelers' Experienced Travel Times for all OD pairs.
104
3000
2500
2000
No hcidet
shcidert- noVMS
0
E 1000 500 -
0:~
Figure 5-7: Impact of Incident. Frequency of Experienced Travel Times for all OD pairs
"1, 2, 3, 4 - 5, 6, 7".
Some important statistics regarding all the OD pairs in the presence of the incident are:
*
8409 vehicles completed their trips as a result of the incident, while 8702 trips were
completed during the no-incident case.
*
The average travel times for travelers who finished their trips between 7:00 AM and
10:00 AM is 2313 seconds.
*
The average travel times based on the departure time interval, aggregated for all OD
pairs, is shown in Figure 5-7a. From this figure, it is clear that for all the OD pairs,
travelers who departed after 7:45 AM experienced delay, as delay propagated until
the end of the interval. The average travel time corresponding to the time interval
8:00 AM - 9:00 AM increased from 2000 seconds to 3000 seconds (approx.).
105
3500
39 0004
i nNoncider :
7".'
iue58
increase bydt647 seWdaSvdntfo
Figure 5-a: mato nietverage
5A1.3mIncident with VMS
7"-
Travel
Time for allth OD pairs
(Diniersoiion trey)dinScin
"nteaayi 1, 2, 3, 4 - 5, 6, 7
.. ied h
aAScomarison[
The average wit
travelthetime
fr lltheconprdictionsVMrestteisrs
OD pair
s is in t , 2,he5..
3, 4yes
- 5, 6,te 7
c no-inidnt
foluna [48]chsonclddtash:rdcieVSsraeisrsl
ntebs
newr pefrmne
networ
esult bot
ne otineterm
n em of the incdene23 fwe
numbernt
offwe travelerswh
completetheir
their tripswihntepro
travelerswh
complete
an The average travel timefoHeafurthe reprs hattinsanaeous VMS
strategies5,might
cause travelers to experience longer travel times, since these strategies do not take into
account future network conditions and may lead to overreaction. This effect may be
106
vrl
avoided by using the predictive VMS scenario with a consistent guidance strategy.
Therefore, this section analyzes the impact due to a predictive VMS strategy only.
Laeations of VMS
2
Tacomc.Snate
Loction of Icdent
Time: 7:45 AM to 9:50 AM
Severity: 80 %capacity reduction
Pkwy
$praia BroOk Pkwy
Saw Mill Rier
Rt 100
fffh a~e
s nLve Pw
Cross County Pkwy
67
Figure 5-8: Location of VMS Message
The VMS strategy in our analysis is constrained by the fact that the location of
VMS on the network is fixed and only these fixed VMS can be used for the diversion of
traffic flow. For all OD pairs (1, 2, 3, 4 - 5, 6, 7) in the analysis, only two VMS, as
shown in Figure 5-8, can be used for the diversion of traffic flow. As is clear from the
locations of these VMS, no guidance can be provided to flow originating from nodes 1
and 2. Only flow for OD pairs "3, 4 - 5, 6, 7" can benefit from these VMS messages.
Response of drivers from nodes 3 and 4 to the incident and VMS will primarily affect
travel times for drivers originating from nodes 1 and 2. This was one more reason why
107
these nodes were aggregated for the incident and VMS analysis. The VMS chosen for
this purpose are link-VMS, which provide information on the travel times of different
links for alternative paths. Travelers update their route choice decision based on updated
links travel times provided by the link-VMS. Link-VMS provide travel time information
of all the links from the upstream of the incident location to the origin of travelers' trip,
since only these links will have higher travel times due to increased congestion caused by
the incident. In this case, link-VMS would be more appropriate since the incident
location is close to the origin nodes and only a few links upstream of the incident location
will be affected by the congestion due to the incident.
5.1.3.1 Impact of VMS Information Strategy
Several iterations were performed to achieve consistency in the prediction and the best
guidance was chosen based on criteria such as the number of finished trips, the average
travel time during the planning horizon and the distribution of the travel times
experienced by travelers. The experienced travel times of travelers plotted as a function
of their departure times is shown in Figure 5-9. Figure 5-9 shows that after using VMS
maximum travel time has decreased from 4500 seconds (incident) to about 4000 seconds.
Figure 5-10 shows the frequency of travel times for aggregated OD pairs (1, 2, 3, 4 - 5, 6,
7). Frequency analysis reveals that in the presence of VMS, many travelers have
experienced reduced delay and many have shifted from higher travel times (3000 - 4500
seconds) to lower travel times.
108
5000
40DO
3500
1500
- 3000
S1500
-
500
60
6:15 6:30
6:45
7:00
7:15
7:30 7:45 8:00
8:15 8:30
8:45
9:00
9:15
9,30
945 10:00
Deparlue Time
Figure 5-9: Impact of VMS. Travelers Experienced Travel Times for all OD pairs.
3500
3000
No;7i en
Z 2500
S2000
iiicdr
a Incidert -no VMS
0
a Incident - VMS
E 1000
z
500
0
4
1000- 1500- 2000- 2500- 3000- 3500- 4000- 45001000 1500 2000 2500 3000 3500 4000 4500 5000
<=
Travel Time (seconds)
Figure 5-10: Impact of VMS. Frequency of Experienced Travel Times for all OD pairs
"1, 2, 3, 4 - 5, 6, 7".
109
-~1
Again, some important statistics regarding all OD pairs under the VMS scenario are
illustrated below:
*
8438 vehicles completed their trips, while 8702 vehicles completed their trips during
the no-incident case and 8409 completed their trips during incident.
9
The average travel time of travelers who finished their trips between 7:00 AM to
10:00 AM is 2084 seconds.
9
The average travel times based on the departure time interval, aggregated for all OD
pairs, is shown in Figure 5-11. Figure 5-11 clearly shows that average travel times in
the VMS scenario decreased compared to the incident scenario. Major reduction in
the delay can be seen during the incident interval (7:45 AM - 8:50 AM). During other
intervals, savings in travel times are marginal.
3500
3000
2500
m No Incident
1 Incident -no VMS
DIncident -Wth VMS
2000
1500~
'1000
CD
S 0
Departure The interval
Figure 5-11: Impact of VMS. Average Travel Time for all OD pairs "1, 2, 3, 4 5, 6, 7".
110
5.1.4 Comparison of Different Cases: No Incident, Incident
without VMS and Incident with VMS
This section compares the results obtained from different scenarios based on the
aggregate statistics, the frequency of trip travel times and the departure time intervals.
5.1.4.1 Comparison Based on Aggregate Statistics
Table 5-1 summarizes the number of completed trips and the average travel times for
different aggregated OD pairs. Further, the percentage change from the no-incident case
is indicated in brackets.
As can be observed from Table 5-1, as a result of the incident, 293 fewer trips
were completed for all OD pairs. Total number of completed trips in the presence or in
the absence of VMS is almost the same. However, in the presence of VMS, there are
savings in travel times. During the incident without VMS, the average travel time for all
OD pairs increased by 37.2 %. In the presence of VMS during the incident, the increase
in the average travel time of all OD pairs decreased from 37.2 % to 23.6 %. Thus, there is
marginal savings in the average travel time using VMS. The main reason for no
significant savings in travel times is that the incident is close to origins and travelers do
not have more options to take alternative routes. In the presence of the VMS during this
incident on Saw Mill River Parkway, drivers can take Sprain Brook Parkway or Taconic
State Parkway. However, even these parkways have heavy traffic during peak hours.
Therefore, diversion to these alternative paths will not result in significant savings.
Completed Trips
No
Incident - no
Incident - with
Incident
VMS
VMS
8702
8409
8438
1686
2313 (37.2%)
2084 (23.6 %)
(All OD pairs: "1, 2, 3, 4 - 5, 6, 7")
Avg. Travel Time (sec)
(All OD pairs: "1, 2, 3, 4 - 5, 6, 7")
Table 5-1: Comparison based on Aggregate Statistics
111
The impact of VMS messages will be illustrated by a more rigorous analysis in the
following sections.
5.1.4.2 Comparison Based on the Frequency of Trip Travel Times
The comparison based on the frequency of trips within various ranges of travel times for
all OD pairs is summarized below in the Table 5-2.
No Incident
Incident - no VMS
Incident - with VMS
Less than 1000 sec
454
388
376
1000-1500sec
2923
2057
2293
1500-2000sec
3131
807
1243
2000-2500sec
1872
1326
1880
2500 - 3000 sec
321
1442
1610
3000-3500sec
1
1569
793
3500
4000 sec
0
720
226
4000 - 4500 sec
0
96
17
4500 - 5000 sec
0
4
0
-
Table 5-2: Comparison based on the Frequency of Trip Travel Times. For All OD Pairs
"1, 2, 3, 4 - 5, 6, 7".
The comparison reveals that as a result of the incident, a significant number of vehicles
with lower travel times (1500 - 2000 seconds) have shifted to higher travel times (3000 5000 seconds). The impact of the VMS is also obvious from the fact that no vehicles
have reported travel times between 4500 and 5000 seconds, and only very few vehicles
have reported travel times in the range of 4000 - 4500 seconds. The number of vehicles
that experienced travel times between 3000 and 4000 seconds also reduced significantly.
The use of VMS has shifted these vehicles from high travel times to travel times between
1500 and 2500 seconds. Thus, the VMS has been largely beneficial in this regard.
112
5.1.4.3 Comparison Based on Departure Time Interval
The following Table 5-3 compares the average travel times for all OD pairs under each of
the scenarios, as a function of the departure-time interval. The percentage change in the
travel time is also indicated in brackets.
No Incident
Incident - no VMS
Incident - with VMS
Avg. TT (s)
Avg. TT (s)
Avg. TT (s)
<= 7:15
1057
1055 (-0.2%)
1057 (0.0%)
7:15 - 7:30
1271
1268 (-0.3%)
1272 (0.0%)
7:30 - 7:45
1700
1721 (1.2%)
1696 (-0.2%)
7:45 - 8:00
1750
2725 (55.8%)
2099 (20.0%)
8:00 - 8:15
1869
3132 (67.6%)
2426 (29.8%)
8:15 - 8:30
2088
3182 (52.4%)
2923 (40.0%)
8:30 - 8:45
2164
3256 (50.5%)
3027 (39.9%)
8:45 - 9:00
2075
2824 (36.1%)
2696 (29.9%)
9:00 - 9:15
1869
2394 (28.1%)
2350 (25.7%)
9:15 - 9:30
1442
1967 (36.4%)
1664 (15.4%)
9:30 - 9:45
1121
1293 (15.4%)
1334 (19.0%)
9:45 - 10:00
0
0
0
Table 5-3: Comparison based on Departure Time Interval. For All OD Pairs "1, 2, 3, 4 5, 6, 7".
Analysis of the average travel times based on departure time interval also hold
promise for the use of VMS in this situation. As a result of the incident, average travel
times for the travelers departing between 7:45 AM to 8:45 AM (incident duration) have
increased by more than 50 %. After using VMS messages, the average travel times for
these departure time intervals have come down by 30%. The increase in the average
travel times after using VMS is only in the 20 to 30 % range. Vehicles departing between
7:45 AM to 8:15 AM have particularly experienced major savings.
113
Based on the above comparisons, the following conclusions are summarized
below:
" The use of predictive VMS is recommended in this case as it has shifted significant
number of vehicles from higher travel times to lower travel times. Similarly there are
large savings in average travel times of vehicles departing during the incident
duration.
" Overall savings in average travel time in the VMS scenario (2084 seconds, Table 5-1)
is 13.6 % compared to the Incident scenario (2313 seconds, Table 5-1), which is also
an achievement.
" After using VMS messages, the total number of completed trips is almost the same as
the incident scenario, but number of travelers with lower travel times is higher.
" As a result of VMS messages, a large number vehicles that have experienced high
travel times (3500 - 5000 seconds) have shifted to moderate travel times (1500 2500 seconds).
*
The potential of VMS to mitigate network conditions during a particular incident
depends to a large extent on the incident location, its severity and on the alternative
paths that circumvent the incident location. During the incident at this location on the
Saw Mill River Parkway, drivers do not have many alternative routes. This is also one
of the main reasons for marginal savings in overall average travel times.
5.2 Scenario 2
This section presents the impact of a minor incident on the network. The planning OD
and the equilibrium travel times obtained from the no-incident case during the time
period 7:00 AM to 09:00 AM were used as inputs for the scenario evaluation. The
scenario that is discussed in this section is based on the impact of another hypothetical
incident. An incident is introduced in the system at the location shown in Figure 5-12.
The incident is created in the east-bound direction of 1-287 and flows from origins 1, 2, 3,
4, 8 to destinations 9, 10, 11 are aggregated for the purpose of incident and VMS
114
analysis. The time of the incident is from 7:45 AM to 8:15 AM and fifty percent of the
segment capacity is available during the incident.
5.2.1 No Incident
As pointed before, the incident being considered for eastbound traffic is minor.
Therefore for eastbound travelers, only the travel times during 7:00 AM to 9:00 AM will
be considered for analysis. Figure 5-13 presents experienced travel times results for eastbound direction travelers between OD pairs "1, 2, 3, 4, 8 - 9, 10, 11", during the period
of 7:00 AM to 9:00 AM. The average travel time for the travelers who completed their
trips in the interval 7:00 AM to 9:00 AM is 1860 seconds and the number of completed
trips in the interval is 2534. Figure 5-14 illustrates the frequency distribution of the travel
times experienced. As can be seen from these figure, maximum number of travelers
experienced travel times between 1000 - 2500 seconds. Also there are very few travelers
who had travel times less than 1000 seconds or more than 3500 seconds. Figure 5-15
gives the average travel time experienced at different time intervals. It has increased from
about 1100 seconds to 2500 seconds, which occurs at 8:45 AM.
115
0-
-------
Locations *f VMS
------
Lcto
4
fi cident
Time:745 AM to 8:15 AM
Sererily: .50 % CVpa ity reduction
2
Rt 9.
TaconicStatePkwy
184;
Sprain BrookPkwy
Saw Mill RiverPkw
Brox River Pkwy.
River PkWY
Cross Couty Py
Figure 5-12: Network showing Origin Nodes (1,2,3,4,8) and Destination Nodes (9,10,11);
Location of the Incident 2 and Examples of Paths for the OD pairs in dark color.
3500
e 2500
2000
E1500
100D
0
6:00
6:15
6:30
6:45
7:00
7:15
7:30
7:45
8:00
8:15
830
8:45
9:00
Departure Time
Figure 5-13: Travelers' Experienced Travel Times (after convergence) between all OD
Pairs "1, 2, 3, 4, 8 - 9, 10, 11".
116
-
.Yt-~~
-
~
----
-
--
-
700
600
16
500
a-
400
E 200 -
z
100
<=1000
1000 -1500 1500 -2000 2000 - 2500 2500 - 3000 3000 - 3500 3500 -4000
Travel Time (seconds)
Figure 5- 14: Frequency of Experienced Travel Times for all OD Pairs "1, 2, 3, 4, 8 - 9,
10, 11".
3000
9f
C 2500
E
1500
~1000
S500
0
<= 715
7:15 7:30
7:30 - 745
745 - 800
8.00-8:15
815 - 830
Departure Time Interval
Figure 5-15: Average Travel Times between OD Pairs "1, 2, 3, 4, 8 - 9, 10, 11".
117
5.2.2 Incident without VMS
To illustrate the impact of the incident alone, the incident (Figure 5-12) was introduced in
the system without VMS and the network performance was analyzed assuming that
travelers follow their habitual routes. The resulting plot of the travel times experienced by
the travelers based on their departure times is shown in Figure 5-16. Since the incident is
from 7:45 AM to 8:15 AM, only the impact between 7:00 AM to 9:00 AM has been
analyzed. In order to illustrate the impact of the incident and as a reference for future
comparisons, the frequency of the experienced travel times is further analyzed based on
the aggregated OD pairs, as done in the previous section. As a result of the incident, the
maximum travel time has increased from 3000 seconds to 3500 seconds, which is not a
significant delay. Figures 5-17 shows the frequency of the travel times for aggregated OD
pairs. Figure 5-17 clearly depicts that as a result of the incident, some of the travelers
experienced high travel time of 3000 - 3500 seconds. Number of travelers having travel
times in the range 1000 - 2500 seconds has reduced significantly and number of travelers
with travel times 2500 - 3000 seconds has increased significantly.
4000
0
2500
M20
1500
0
6:0
6:15
6:30
6:45
7:00
7:15
7:30
7:45
8:00
8:15
830
8:45
9:D0
Deprure Tim
Figure 5-16: Impact of Incident. Travelers' Experienced Travel Times for all OD pairs.
118
700d
500 E 00
z
0
<=1000
1000-1500 1500-2000 2000- 2500 2500 - 3000 3000 -350
350 - 4000
TraveIlime (seconds)
Figure 5-17: Impact of Incident. Frequency of Experienced Travel Times for all OD
pairs "1, 2, 3, 4, 8 -9, 10, 11".
Some important statistics regarding all the OD pairs in the presence of the incident are:
0
2457 vehicles completed their trips as a result of the incident, while 2534 trips were
completed during the no-incident case.
*
The average travel time of the travelers who finished their trips between 7:00 AM and
9:00 AM is 2070 seconds.
*
The average travel times based on the departure time interval, aggregated for all OD
pairs, is shown in Figure 5-18. From this figure, it is clear that for all the OD pairs,
travelers that departed between 7:45 AM to 8:15 AM have experienced marginal
delay.
119
3500
3000
2500
2000
aNo incidert
kiident -no VMS
1500
1000
500
0
<7:15
715-730
730-7:45
7:45-8:00
8.00-8:15
8:15-&30
Figure 5-18: Impact of Incident. Average Travel Time for all OD pairs "1, 2, 3, 4, 8 - 9,
10, 11".
5.2.3 Incident with VMS (Diversion Strategy)
Again, the VMS strategy in our analysis is constrained by the fact that the location of
VMS on the network is fixed and only these fixed VMS can be used for the diversion of
traffic flow. For all OD pairs (1, 2, 3, 4, 8 - 9, 10, 11) in the analysis, only three VMS, as
shown in Figure 5-12, can be used for the diversion of traffic flow. As is clear from the
locations of these VMS, no guidance can be provided to flow originating from node 8.
Only flow for OD pairs "1, 2, 3, 4 - 9, 10, 11" can benefit from these VMS messages.
The VMS chosen for this purpose are link-VMS, which provide information on the travel
times of different links for alternative paths.
5.2.3.1 Impact of VMS Information Strategy
Several iterations were performed to achieve consistency in the prediction and the best
guidance was chosen based on criteria such as the number of finished trips, the average
120
travel time during the planning horizon and the distribution of the travel times
experienced by travelers. The experienced travel times of travelers plotted as a function
of their departure times is shown in Figure 5-19. It can be seen that the maximum travel
time is again close to 3000 seconds, which is also no-incident case maximum travel time.
Figure 5-20 shows the frequency of travel times for all OD pairs (1, 2, 3, 4, 8 - 9, 10, 11).
It is clear from the Figure 5-20 that number of travelers with travel times more that 3000
seconds has reduced and with travel times of 2000 - 2500 seconds have doubled
4000 -
E
1500
S1000
I-
500
0
6:00
6:15 6:30
6:45
7:00
7:15 730 7:45 800
Depautre Tine
8:15
8:30
8:45
9.00
Figure 5-19: Impact of VMS. Travelers Experienced Travel Times for all OD pairs.
121
700
600
500ZD
hddent - noVG
300-
0 hident - wi VMS
200
z
100.
04
<=1000
1000- 1500 1500 -2000 2000 -2500 2500 - 3000 300
Travel Time (secnds)
3500 3500 -4000
Figure 5-20: Impact of VMS. Frequency of Experienced Travel Times for all OD pairs
"1, 2, 3, 4, 8 - 9, 10, 11".
Again, some of the important statistics regarding all OD pairs under the VMS scenario
are illustrated below:
*
2477 vehicles completed their trips, while 2457 vehicles completed their trips during
the incident and 2534 completed during the no-incident case.
*
The average travel time for the travelers who finished their trips between 7:00 AM
and 9:00 AM is 1942 seconds.
*
The average travel times based on the departure time interval for aggregated OD pairs
is shown in Figures 5-21. Figure 5-21 clearly shows that average travel time in the
VMS scenario decreased compared to the incident scenario. In the presence of VMS
information, average travel times are closer to the no-incident case average travel
times. The effect of VMS is more pronounced between 7:45 and 8:00.
122
3500
U
0
U
S
U
3000
2500
S
nNo Incident
QIlncident -no VMS
'ancident -with VMS
E 2000
0
1500
L
I0
1000
U
L
S
500
0
7:30 -7:45
745 -8:00
8:00- 8:15
8:15- 8:30
Departure Tiff Interval
Figure 5-21: Impact of VMS. Average Travel Time for all OD pairs "1, 2, 3, 4, 8 - 9, 10,
11".
5.2.4 Comparison of Different Cases: No Incident, Incident
without VMS and Incident with VMS
This section compares the results obtained from different scenarios, based on the
aggregate statistics, the frequency of trip travel times and the departure time intervals.
5.2.4.1 Comparison Based on Aggregate Statistics
Table 5-4 summarizes the number of completed trips and the average travel time for
different aggregated OD pairs. Further, the percentage change from the no-incident case
is indicated in brackets.
As can be observed in Table 5-4, with the presence of VMS during the incident,
the total number of completed trips is almost same as those during the incident. 77 fewer
trips were completed for all OD pairs when VMS was not used in the presence of the
incident. During the incident without VMS, the average travel time for all OD pairs
123
increased by 11.3 %. However, in the presence of VMS during the incident, the increase
in the average travel time for all OD pairs decreased from 11.3 % to 4.4 %. Thus, there
are savings in the average travel time by using VMS. However, the average travel time
after implementing the VMS (1942 seconds) is still larger than the average travel time
without the incident (1860 seconds).
Completed Trips
No
Incident - no
Incident - with
Incident
VMS
VMS
2534
2457
2477
1860
2070 (11.3 %)
1942 (4.4 %)
(All OD pairs: "1, 2, 3, 4, 8 - 9, 10, 11")
Avg. Travel Time (sec)
(All OD pairs: "1, 2, 3, 4, 8 - 9, 10, 11")
Table 5-4: Comparison based on Aggregate Statistics
The impact of VMS messages will be illustrated by a more rigorous analysis in the
following sections.
5.2.4.2 Comparison Based on the Frequency of Trip Travel Times
The comparison based on the frequency of trips within various ranges of travel times for
all OD pairs is summarized below in Table 5-5.
No Incident
Incident - no VMS
Incident - with VMS
Less than 1000 sec
202
200
202
1000 - 1500 sec
595
569
577
1500 - 2000 sec
632
471
516
2000 - 2500 sec
665
327
599
2500
3000 sec
430
570
426
3000-3500sec
10
300
157
3500-4000sec
0
20
0
-
Table 5-5: Comparison based on the Frequency of Trip Travel Times. For All OD Pairs
"1, 2, 3, 4, 8 - 9, 10, 11".
124
The comparison reveals that as a result of the incident, a significant number of vehicles
with travel times 1500 - 2500 seconds have shifted to higher travel times (2500 - 4000
seconds). The impact of the VMS is also obvious from the fact that no vehicles reported
travel times between 3500 to 4000 seconds. The number of vehicles who experienced
travel times between 2500 and 3500 seconds also reduced significantly. The use of VMS
has shifted these vehicles from high travel times to travel times between 1500 and 2500
seconds. Thus, the VMS has been largely beneficial in this regard.
5.2.4.3
Comparison Based on Departure Time Interval
The following table compares the average travel times for all OD pairs under each of the
scenarios, as a function of the departure-time interval. The percentage change in the
travel time is also indicated in brackets.
No Incident
Incident - no VMS
Incident - with VMS
Avg. TT (s)
Avg. TT (s)
Avg. TT (s)
<= 7:15
1125
1122 (-0.3%)
1113 (-1.1%)
7:15 - 7:30
1584
1584 (0.0%)
1606 (1.4%)
7:30 - 7:45
1974
2284 (15.7%)
2093 (6.1%)
7:45 - 8:00
2333
2959 (26.9%)
2506 (7.4%)
8:00 - 8:15
2530
2825 (11.7%)
2774 (9.6%)
8:15 - 8:30
2311
2339 (1.2%)
2273 (-1.7%)
8:30 - 8:45
0
0
0
Table 5-6: Comparison based on Departure Time Interval. For All OD Pairs "1, 2, 3, 4, 8
-9, 10, 11".
The average travel times in a departure time interval also hold promise for the use
of VMS in this situation. As a result of the incident, average travel times of the travelers
departing between 7:45 AM and 8:00 AM increased by more than 25 %. Although
incident starts at 7:45 AM, vehicles departing between 7:30 AM and 7:45 AM have also
125
experienced an increase of 15 % in the average travel time. This increase is due to the
fact that the incident had already begun by the time the vehicles departing between 7:30
AM and 7:45 AM reached the incident location. After using VMS messages, the average
travel time for this departure time interval has almost halved. The increase in the average
travel times after using VMS is only in the 5 to 10 % range. Vehicles departing between
7:30 AM to 8:15 AM have particularly experienced major savings.
Based on the above comparisons, some conclusions are summarized below:
" The use of predictive VMS is again recommended in this case as it has resulted in the
shift of vehicles from higher travel times to lower travel times. Similarly, vehicles
departing between 7:45 AM and 8:15 AM have significant savings in average travel
times.
" After using VMS messages, the total number of completed trips is almost the same as
those in the Incident scenario. However, average travel times have decreased after
using VMS and number of vehicles with high travel times (3000 - 4000 seconds)
during incident has shifted to lower travel times (1500 - 2500 seconds).
*
The potential of VMS to mitigate network conditions during a particular incident
depends to a large extent on the incident location, its severity and on the alternative
paths that circumvent the incident location. In this case, vehicles have option to divert
from 1-287 to arterial roads at the upstream of the incident and they can later merge at
1-287 at the downstream end of the incident. This might explain that average travel
times after using VMS are close to no-incident case average travel times.
5.3 Summary
This chapter illustrated the potential of the planning tool (Sundaram [48]) in evaluating a
VMS case study in the Lower Westchester County, New York. The Incident and VMS
scenarios were evaluated based on hypothetical incidents on the LWC network.
DynaMIT-P was used to evaluate predictive information through the VMS. The planning
tool developed by Sundaram [48] was found to be extremely useful in analyzing various
scenarios and capturing the relevant details. The case study was intended to give a sense
of the potential of DynaMIT-P and illustrate the benefits of deploying such planning
126
tools. The next chapter focuses on areas where further research may be employed to
improve the functionality of the planning tool.
127
128
Chapter 6
Conclusion
Dynamic Traffic Assignments
(DTA) system can be used for dynamic traffic
management purposes, like reducing delay on major highways, improving safety,
efficiency and capacity of transportation network, and improving wide-area emergency
responses through information sharing and coordination. DynaMIT-P is a DTA-based
planning tool that operates in an offline mode, in which traffic managers investigate and
prepare standard response strategies to different traffic and incident scenarios. The
objective of this thesis was to apply DynaMIT's DTA capabilities to the Lower
Westchester County (LWC) ITS subsystem and to use calibrated system to perform
illustrative analyses of incident response strategies. First, DynaMIT-P has been calibrated
in order to estimate traffic conditions in the LWC network with precision sufficient for
ITS purposes. The results of a case study, focusing on the evaluation of the diversion
response strategy in case of an incident on the LWC network, illustrate the functionality
and potential of the system.
6.1 Thesis Contributions
This thesis used the simulation-based Dynamic Traffic Assignment (DTA) planning tool
for short-term applications developed by Sundaram [48]. This tool overcomes the
drawbacks of the traditional planning tools, which are static in nature and are not
adequate to evaluate planning strategies that involve explicit modeling of traffic
dynamics and traveler behavior. To evaluate the diversion strategies during incidents it is
necessary to capture traffic dynamics and traveler behavior. Therefore, this planning tool
was able to capture the different scenarios more accurately. Moreover, this tool also
models day-to-day behavior, within-day behavior of travelers, and response of travelers
to information.
129
The planning tool was first calibrated from the limited data available for the
network. Some assumptions regarding the sensor data were made, as sensor data from the
same date was not available for calibration purpose. As a result, a thorough analysis of
the traffic sensor data available was done and sensors that were not able to capture the
traffic dynamics were removed from the analysis. The no-incident scenario was obtained
first during the calibration process, which not only established the equilibrium travel time
conditions in the network, but also provided an OD matrix that best reflected the
observed sensor counts. The no-incident case captured the day-to-day behavior of
travelers in the network.
The impact of two hypothetical incidents, in the absence of VMS messages, was
analyzed by considering that travelers will follow their habitual paths. One incident was
for 65 minutes on Saw Mill River Parkway and twenty percent of the segment capacity
was available. The results show that the impact of the incident is severe as many vehicles
reported travel times more than 4000 seconds, which was not observed during normal
traffic conditions. The generation of the consistent predictive information was used for
the VMS information systems. The result of the VMS scenario shows that predictive
VMS is effective because it not only causes a large number of travelers to experience
lower travel times, but average travel times experienced during different departure
intervals also decrease. The use of VMS is recommended if the incident occurs at the
same location since there is saving in travel times and more number of travelers is able to
shift to lower travel times.
6.2 Scope of Future Research
This thesis applied the simulation-based DTA
system for short-term planning
applications. There are several issues that could be pursued in future.
The calibration of the DTA system was the most challenging part since demand
calibration and supply calibration were done separately. A natural extension would be to
focus on the integrated calibration of both the demand and supply simulator modules in
the DTA system. During different steps of calibration process, it was clearly observed
130
that better quality of field data is crucial for better estimates of demand and supply
parameters.
It is very important to understand and model traveler behavior in response to
information. Although many studies have been done in this direction, there is still
significant scope for improvement to the existing models. Incorporation of traveler
behavior models that closely mimic traveler response to information will enhance the
importance of planning tools.
131
132
Appendix A
DynaMIT-P Input Files
Following are the brief details of the input files required by DynaMIT-P system:
(1) The network file contains the description of the network under study like nodes,
links, segments, lanes and sensors. It also has the connectivity between different
lanes in the network. Changing geometry within links is modeled by dividing the
links into smaller segments.
(2) The supply parameter file contains the segment-specific details of parameters
from the speed-density relationship, and segment capacities. These parameters are
used by the supply simulator for simulating the movement of vehicles on the
network.
(3) The demandfile specifies the time-varying origin-destination flows for each OD
pair.
(4) The link travel times file is used for the assignment of routes of vehicles.
(5) The socio-economic file has information regarding various socio-economic
characteristics of drivers like trip purpose, value of time, access to ATIS etc.
These characteristics are used while generating the population of drivers and are
defined for each origin-destination pair in the demand file.
(6) The behavior parameters file has values of different coefficients in habitual, pretrip and en-route route choice models.
(7) The sensor count file has the time-varying traffic sensor counts that are used
during OD estimation process to map OD flows to sensor counts.
(8) In addition to the inputs discussed above, DynaMIT-P's demand simulator
requires error covariance and autoregressive matrices that are used during OD
estimation. These matrices result from a process of calibration.
133
Appendix B
The vms.dat File
Information pertaining to the presence of VMS and HAR devices in the network is
contained within the vms . da t file. As discussed earlier, DynaMIT has been enhanced to
allow for two types of VMS: one that provides travel time guidance as it pertains to
single links, and another that provides travel time guidance as applicable to paths
(subpaths). Furthermore, HAR devices are implemented as groups of VMS that cover the
area for which traffic information can be provided through the HAR devices.
To
illustrate this modeling convention we will once again use the same simple network
(Figure B-1) as an example. A HAR whose range covers links 1, 3, and 4 can be modeled
as three VMS (one located in each of the following links: 1, 3, and 4).
3
VMS
5
AR range
10
6
7
Direction of flow
8
11
1
Figure B-1. Example to Illustrate link-VMS and path-VMS
In this section the structure of the input file that models VMS and HAR devices
will be introduced. Furthermore, a sample file will illustrate the structure of the vms.dat
file and its use in modeling realistic situations. The structure of the file is as follows. The
first line contains two numbers. The first number provides the number of link-VMS and
the second number provides the number of path-VMS devices (we assume that HAR
devices have already been translated into equivalent VMS devices).
The syntax of the first line of the file then takes the following generic form:
#linkVMS #pathVMS
134
The remainder of the file is composed by descriptions of each device, in the following
order (assuming k link-VMS6 and 1 path-VMS7):
Link VMS 1 description
Link VMS k description
Path VMS 1 description
Path VMS 1 description
The structure of each device description follows. A link VMS description has the
following structure:
Link on which theVMSis
#_oflinkstheVMSprovidesinfofor
listoflinksforwhichinfo
{
isprovided }
For example, the following link-VMS definition:
1
2 { 2 3}
implies that the particular VMS is located in link 1 (indicated in the first line) and
provides information about 2 links (first number of second line). The links for which
information is provided are links 2 and 3 (two numbers within brackets in second line).
A path VMS description has the following structure:
Link on which theVMSis ...
number_ofpathsforwhichinformation_is_provided
#_oflinksonfirst-path_forwhichinfois-provided
6 We assume that link-HAR have already been translated into equivalent link-VMS.
7 We assume that path-HAR have already been translated into equivalent path-VMS.
135
{
linksthatmakeup-path }
#_oflinkson last-path forwhichinfoisprovided
{
linksthatmakeup-path }
For example, the following path-VMS definition:
4 2
1 {5}
3 {6
7
8}
implies that the particular VMS is located in link 4 (indicated by the first number in the
first line) and provides information about 2 paths (second number of first line). Each
subsequent line provides information for each of the paths for which information is
provided. The first path is composed of one link (first number of second line) and is
composed of link 5 (number within brackets). Similarly, the second path is composed of
three links (first number of last line) and these links are 6, 7 and 8 (the three numbers
within the brackets at the remainder of the line).
The sample file below elucidates the structure of the file for a network with two VMS
giving link (travel time) information and one VMS providing path (travel time
information):
2 1
9
1 { 10
1
2 {2 3}
4 2
1 {5}
3 {6
7
8}
This entire file can be thought of as consisting of three sections, the first section having
just a single line. The first line indicates that there are two VMS giving link (travel time)
information and one VMS giving path (travel time) information. Figure B-2 shows the
136
file with explanatory information. This figure should be read in conjunction with the
following paragraphs.
The next section that follows consists of (two times the number of link VMS) lines
containing information about the link-guidance VMS. Thus, the second and third lines
indicate that the first link VMS is on link 9 and that it provides information for 1 linklink number 10. Similarly, the fourth and fifth lines indicate that the second link VMS is
on link 1 and provides information for 2 links-link 2 and link 3.
The third and last section contains information about the path-VMS. The first line in this
section denotes the fact that the path guidance VMS is on link 4 and provides guidance
for 2 paths. Each such line (one for every path-VMS) will be followed by as many lines
as the number of paths for which guidance is being provided. The second line implies that
the first of these two paths has just 1 link-link 5, and the third line implies that the
second path has three links-links 6, 7 and 8 in that order.
#link-VMS
#path-VMS
2 1
VMS is on link 9
VMS provides
information for 1
link
VMS provides
information for
VMS provides information for link 8
VMS is on link 1
VMS provides information for links 2 and 3
path-VMS is on
link 4
4 2
first pankhas 1
1
VMS provides information for 2 paths
path is composed of link 1
5
path is composed of links 6, 9, and 7
second path has 3
links
Figure B-2. Sample vms.dat File with Explanatory Information
137
Appendix C
Analysis
Data obtained from
of Sensor
NYSDOT for LWC Network
Key Points:
*
All sensors data obtained have HOURLY count.
" All sensors data are from WEEK DAYS.
*
Most of the sensors data are for almost* 3 days.
*
Each sensor means a pair of sensors in each direction (out of a pair of NB-SB or EBWB) of traffic. Like, if there are 6 sensors data on 1-87. It means there are actually 12
(6x2: 1 for North-Bound and 1 for South-Bound) sensors data.
All sensors in the LWC network are divided into 2 categories: (a) Freeway Sensors, and
(b) Parkway Sensors
(a)
Freeway Sensors
There are 4 freeways in the network. These are 1-87, 1-95, 1-287 and 1-684.
(i) 1-87/NYS Thruway
7 sensor data are obtained for this freeway. 4 out of 7 sensors have data recorded in
September 2002, 2 sensors have data from November 2000 and 1 sensor data is from
November 2001.
138
3 out of 4 sensors data of 2002 have almost* 3 days of data starting from September 17,
2002. Remaining sensor also have almost 3 days data but starting from September 16,
2002.
2 sensors data from November 2000 have 3 days of data starting from November 6, 2000.
1 sensor data from November 2001 has almost 3 days of data starting from November 7,
2001.
(ii) 1-95
1-95 has 9 sensors data. 3 out of 9 sensors have data from September 2002 and rest 6 have
data from November 2000.
3 sensors data of 2002 have almost 3 days of data starting from September 17, 2002. 6
sensors data of 2000 have 3 days of data starting from November 6, 2000.
(iii) 1-287/Cross Westchester Express
1-287 has 7 sensors data. 1 out of 7 sensors data is from September 2002, 4 sensors have
data from November 2000 and rest 2 from November 2001.
1 sensor data of 2002 has almost 3 days of data starting from September 17, 2002. 4
sensors data of 2000 have 3 days of data starting from November 6, 2000. 2 sensors data
from 2001 have almost 3 days of data starting from November 7, 2001.
(iv) 1-684
* Word "almost" has been used deliberately to emphasis the fact that data for some sensors are less
than 3
days but more than 2 day. Also for some sensors "starting time" and "ending time" for data are different
even if they are for 3 days or for "almost" 3 days.
139
1-684 has 7 sensors data. 6 sensors data are for almost 3 days starting from October 16,
2000 and 1 sensor data from 2001 has almost 3 days of data starting from August 27,
2001.
(b) Parkway Sensors
There are 6 parkways in the network. These are: (i) Saw Mill River Parkway (ii) Bear
Mountain Parkway (iii) Cross Country Parkway (iv) Taconic State Parkway (v) Sprain
Brook Parkway, and (vi) Hutchinson River Parkway.
(i) Saw Mill River Parkway
This parkway has 7 sensors data. 3 out of 6 sensors have data from July-August 2000, 3
have data from October 2000 and 1 from September 2001.
3 sensors data of October 2000 have almost 3 days of data starting from October 23,
2000. 1 sensor data from 2001 has 3 days of data starting from September 10, 2001. 1
sensor data of July-August 2000 has 1 week of data (including Saturday and Sunday)
starting from July 31. Rest 2 sensors of July-August 2000 have 3 days of data starting
from July 31.
(ii) Bear Mountain Parkway
This parkway has 2 sensors data. 1 sensor has data from June 2000 and other has data
from April-May 2002.
1 sensor data of April-May 2002 has almost 3 days of data starting from April 29, 2002. 1
sensor data of June 2000 has 3 days of data starting from June 12, 2000.
140
(iii) Cross Country Parkway
This parkway has one sensor data from October 2000. This sensor has almost 3 days of
data starting from October 16, 2000.
(iv) Taconic State Parkway
This parkway has 9 sensors data from June 2000. All sensors have almost 3 days or 3
days of data starting from June 12, 2000 (except for 2 SB sensors out of total 18 (=9x2)
sensors). These 2 SB sensors have almost 3 days of data for the same day of week but on
different date (i.e. June 26, 2000).
(v) Sprain Brook Parkway
This parkway has 6 sensors data. 3 out of 6 sensors data are from July-August 2000, 2
sensors data are from October 2000 and rest 1 is from September 2001.
2 sensors data of October 2000 have almost 3 days of data starting from October 23,
2000. 3 sensors data of July-August 2000 have 3 days of data starting from July 31, 2000.
1 sensor data from September 2001 has 3 days of data starting from September 10, 2001.
(vi) Hutchinson River Parkway
This parkway has 11 sensor data. 4 out of 11 sensors data are from May 2002, 1 sensor
data is from October 2000 and rest 6 have sensors data from June 2000.
4 sensors data of May 2002 have almost 3 days of data starting from May 14, 2002. 1
sensor data of October 2000 has almost 3 days of data starting from October 16, 2000. 6
sensors data of June 2000 have almost 3 days of data starting from June 12, 2000 (except
for 1 NB and 1 SB sensors). This 1 NB sensor has data for almost 3 days but starting
141
from June 13, 2000 and 1 SB sensor has data for almost 3 days but taken 2 weeks later
(i.e. June 26, 2000).
Summary: Initially, there were 120 sensors data received from NYSDOT. These include
52 sensors from freeways and 68 sensors from parkways. 36 of 52 freeway sensors were
from year 2000 and rest 16 were from 2002. 58 of 68 parkway sensors were from 2000
and rest 10 from 2002.
After removing 32 sensors that were not in our network area, 88 sensors were left in the
focused network. These included 42 sensors from freeways and 46 from parkways. 26 of
42 freeway sensors were from 2000 and rest 16 from 2002. 38 of 46 parkway sensors
were from 2000 and rest 8 from 2002.
Later 88 sensors were supplemented by 12 more sensors from year 2001 to make it to
total of 100 sensors. 8 of 12 sensors were from freeways and rest 4 from parkways.
After calibration results it was observed that 42 out of 100 sensors are not able to capture
the real traffic conditions in the field, and as a result these 42 sensors were removed from
the network making it to total of 58 sensors in the network. This removal of 42 sensors
was done step by step as calibration process was progressing. 32 of 58 sensors are from
freeways and 26 from parkways. 20 of 32 freeway sensors are from 2000, 1 from 2001
and 11 from 2002. 22 of 26 parkway sensors are from 2000, 1 from 2001 and 3 from
2002. In terms of year, 44 sensors are from 2000, 2 from 2001 and rest 14 from 2002.
142
143
Bibliography
[1] Abdulhai, B. and T. Ma. GENOSIM: A Genetic Algorithm-based Optimization
Approach and Generic Tool for the Calibration of Traffic Microscopic Simulation
Parameters.TransportationResearch Board. 2002
[2] Al-Deek, H. and Kanafani, A. (1993). Modeling the Benefits of Advanced Traveler
Information Systems in Corridors with Incidents. TransportationResearch, 1993.
[3] Arnott, R., de Palma, A. and Lindsey, R. (1991) Does Information to Drivers Reduce
Traffic Congestion? TransportationResearch, 1991.
[4] Ashok, K. Estimation and Prediction of Time-Dependent Origin-Destination Flows.
Ph.D. thesis, Department of Civil and Environmental Engineering, Massachusetts
Institute of Technology, Cambridge, MA 02141.
[5] Ashok, K. and Ben-Akiva, M.E. Estimation and Prediction of Time-Dependent
Origin-Destination Flows with a Stochastic Mapping to Path Flows and Link Flows.
Resubmitted to TransportationScience, 1999.
[6] Balakrishna, R.: 2002, Calibrationof the Demand Simulator in a Dynamic Traffic
Assignment Systems, Master's
thesis, Department of Civil
and Environmental
Engineering, Massachusetts Institute of Technology, Cambridge, MA.
[7] Bell, M.G.H. A Game Theory Approach to Measuring the Performance Reliability of
Transportation Network, TransportationResearch, Part B, 1999.
[8] Ben-Akiva, M.E. and Lerman, S. Discrete Choice Analysis: Theory and Application
to Travel Demand. MIT Press, Cambridge, MA USA 1985.
144
[9] Ben-Akiva, M.E., Bierlaire, M., Koutsopoulos, H.N. and Mishalani, R. Real Time
Simulation of Traffic Demand-Supply Interactions within DynaMIT. Technical report,
ROSO-DMA-EPFL, 2000.
[10] Ben-Akiva, M., de Palma, A. and Kaysi, I. Dynamic Network Models and Driver
Information Systems. TransportationResearch, 1991.
[11] Bierlaire, M., Mishalini, R., and Ben-Akiva, M.E. General Framework for Dynamic
Demand Simulation. Submitted to TransportationResearch B, 2000.
[12] Bonsall, P. W. and Merrall A. C. (1997) Analyzing and Modeling the Influence of
Roadside Variable Message Display on Drivers' Route Choice. In Hensher, D. A., King,
J., Oum, T. H.(Eds.), World Transport Research: Proceedings of the 7th World
Conference on TransportResearch, vol. 1: Traveler Behavior. Elsevier, Oxford, 1997.
[13] Brandriss, J.J. Estimation of Origin-Destination Flows for Dynamic Traffic
Assignment. Master's thesis, Department of Civil and Environmental Engineering,
Massachusetts Institute of Technology, Cambridge, MA, 2001.
[14] Cascetta, E. Estimation of Trip Matrices from Traffic Counts and Survey Data: A
Generalized Least Square Estimator. TransportationResearch, 1984.
[15] Cascetta, E. and Cantarella, G. A Day-to-Day and Within-Day Dynamic Stochastic
Assignment Model. TransportationResearch, 1991.
[16] Cascetta, E., Inaudi, D. and Marquis, G. Dynamic Estimators of Origin-Destination
Matrices using Traffic Counts. TransportationScience, 1993.
[17] Cascetta, E., Nuzzolo, A. and Russo, F. and Vitetta, A. A modified Logit Route
Choice Model Overcoming Path Overlapping Problem:
Specification and Some
Calibration Results for Inter-urban Networks. Transportationand Traffic Theory, 1997.
145
[18] Chatterjee, K., Hounsell, N. B., Firmin, P. E. and Bonsall, P. W. (2000) Driver
Response to Variable Message Sign Information in London. Transportation Research,
2000.
[19] Dahlgren, J. High-Occupancy / Toll Lanes: Where Should they be Implemented?
TransportationResearch, 2002.
[20] Darda, D.: 2002, Joint Calibration of a Microscopic Traffic Simulator and
Estimation of Origin-Destination Flows, Master's thesis, Department of Civil and
Environmental Engineering, Massachusetts Institute of Technology, Cambridge, MA.
[21]
Federal Highway Administration:
1995, User Manual -
CORSIM V1.0,
http://www.fhwa-tsis.com.
[22] Gardes, Y. and May, A. D. (1993) Simulation of IVHS on the Santa Monica
Freeway Corridor using the INTEGRATION model. Phase 2: Preliminary ATIS and
ATMS Experiments. PATH Research Report UCB-ITS-PWP-93-6, 1993.
[23] Hall, R. W. (1993) Non-recurrent Congestion: How Big is the Problem? Are
Traveler Information Systems the Solution? TransportationResearch, 1993.
[24] Hiroshi, W. and Iida Y. Improvement of Road Network Reliability with Traffic
Management. IFAC Transportation Systems, 603-608, 1994.
[25] Horowitz, J.H. The stability of Stochastic Equilibrium in a Two-Link Transportation
Network. TransportationResearch, 1984.
[26] Institute for Transportation Studies, University of Leeds. SMARTEST: Simulation
Modeling Applied to Road Transport European Scheme Tests, Best Practice Manual,
http://www.its.leeds.ac.uk/smartest, May 1999.
146
[27] Jayakrishnan, R., Cohen, M., Kim, J., Mahmassani, H. S. and Hu, T-Y. (1993) A
Simulation-BasedFrameworkfor the Analysis of Traffic Networks Operating with RealTime Information. PATH Research Report UCB-ITS-PRR-93-25, 1993.
[28] Jha, M., Madanat, S. and Peeta, S. Perception Updating and Day-to-Day Travel
Choice Dynamics in Traffic Networks with Information Provision. Transportation
Research, 1998.
[29] Johnston, R. A. and Ceerla, R. The Effects of New High-Occupancy Vehicle Lanes
on Travel and Emissions. TransportationResearch, 1996.
[30] Koutsopoulos, H. and Yand
Q.:
1996, A Simulation Laboratory for Evaluation of
Dynamic Traffic Management Systems, Department of Civil and Environmental
Engineering, Massachusetts Institute of Technology, Cambridge, MA.
[31] Kunde, K.: 2002, Calibration of the Supply Simulator in a Dynamic Traffic
Assignment System, Master's thesis, Department of Civil and Environmental Engineering,
Massachusetts Institute of Technology, Cambridge, MA.
[32] Kurian, M.: 2000, Calibration of a Microscopic Traffic Simulator, Master's Thesis,
Department of Civil and Environmental Engineering, Massachusetts Institute of
Technology, Cambridge, MA.
[33] Lotan, T. and Koutsopoulos, H.N. Approximate Reasoning Models for Route Choice
Behavior in the Presence of Information. In Daganzo, C. (Ed.), The
12 1h
International
Symposium on Transportationand Traffic Theory (ISTT), Berkeley, CA, 1993.
[34] Mahmassani, H.S. and Chang, G. Experiments with Departure Time Choice
Dynamics of Urban Commuters. TransportationResearch, 1986.
147
[35] Mahmassani, H.S. and Chang, G.L. Dynamic Aspects of Departure-Time Choice in
a
Commuting
System:
Theoretical
Framework
and
Experimental
Analysis.
TransportationResearch Record, 1985.
[36] Mahmassani, H.S., Hu, T-Y. and Jayakrishnan, R. Dynamic Traffic Assignment and
Simulation for Advanced Network Informatics (DYNASMART). In Proceedings of the
2 " InternationalCAPRI Seminar on Urban Traffic Networks, Capri,Italy, 1992.
[37] Massachusetts Institute of Technology. Development of a Deployable Real-Time
Dynamic Traffic Assignment System, Executive Summary: DynaMIT and DynaMIT-P,
version 0.90. Technical Report, Submitted to Oak Ridge National Laboratories, June
2000.
[38] Massachusetts Institute of Technology. DTA System Enhancement and Evaluation
at Traffic Management Center, Task P: Framework for the Use of DynaMIT in
Transportation Planning Applications (DynaMIT-P). Technical report, Submitted to Oak
Ridge National Laboratories, March 2000.
[41] Massachusetts Institute of Technology: 1999, Development of ITS Analysis Toolsfor
Lower Westchester County, Project proposal, Submitted to New York State Department
of Transportation.
[40] McArthur, D. (1995) The PARAMICS-CM (parallel microscopic traffic simulator
for congestion management) behavioral model. In Transportation Planning Methods.
Proceedings of Seminar E held at the 23rd European Transport Forum, University of
Warwick, England, 11-15 September, P412. PTRC Education and Research Services,
London, 1995.
[41] Nicholson, A. and Du. Z. Degradable Transportation Systems: Sensitivity and
Reliability Analysis. TransportationResearch, B, Vol. 31, No. 3, 225-237.
148
[42] Nicholson, A. and Du. Z. Degradable Transportation Systems: An Integrated
Equilibrium Model. TransportationResearch, B, Vol. 31, No. 3, 209-223.
[43] Peeta, S., Ramos, J. L., and Pasupathy, R. (2000) Content of Variable Message Signs
and On-Line Driver Behavior. Submitted for Presentation at the 79th Annual Meeting of
the Transportation Research Board and Publication in Transportation Research Record,
2000.
[44] PTV AG: 2001, VISSIM V3.60 - User Manual, http://www.ptvag.com.
[45] Quadstone Limited: 2003, PARAMICS V4.1 - Modeler Reference Manual,
http://www.paramics-online.com.
[46] Ramming, S. Network Knowledge and Route Choice. Ph.D. thesis, Department of
Civil
and Environmental
Engineering,
Massachusetts
Institute
of Technology,
Cambridge, MA, 2001
[47] Simon, H. A Behavioral Model of Rational Choice. Q.J. Econ, 1955.
[48] Sundaram, S.: 2002, Development of a Dynamic Traffic Assignment System for
Short-Term Planning Applications, Master's
thesis,
Department
of Civil
and
Environmental Engineering, Massachusetts Institute of Technology, Cambridge, MA.
[49] Toledo, T.: 2003, IntegratedDriver Behavior Modeling, Ph.D. thesis, Department of
Civil
and
Environmental
Engineering,
Massachusetts
Institute
of Technology,
Cambridge, MA.
[50] Taylor, N.B. CONTRAM 5: An Enhanced Traffic Assignment Model, TRL
Research Report 249, 1990.
149
[51] Walting, D. and Van Vuren, T. (1993) The Modeling of Dynamic Route Guidance
Systems. TransportationResearch, 1993.
[52] Wardman, P., Bonsall, P. W., and Shires, J. D. (1997) Driver Response to Variable
Message Signs: A Stated Preference Investigation. TransportationResearch, 1997.
[53] Ziliaskopoulos, A.K., Waller, S.T., Li Y., Byram, M. Large Scale Dynamic Traffic
Assignment:
Implementation
Issues and Computational
Analysis.
Submitted for
publicationto the ASCE Journalof TransportationEngineering,October 2000.
150
S2 7
- 22